This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Agent Computing and Multi-Agent Systems 10th Pacific Rim International Conference on Multi-Agents, PRIMA 2007 Bangkok, Thailand, November 21-23, 2007 Revised Papers
13
Series Editors Randy Goebel, University of Alberta, Edmonton, Canada Jörg Siekmann, University of Saarland, Saarbrücken, Germany Wolfgang Wahlster, DFKI and University of Saarland, Saarbrücken, Germany Volume Editors Aditya Ghose School of Computer Science and Software Engineering University of Wollongong Wollongong, NSW, Australia E-mail: [email protected] Guido Governatori National ICT Australia Queensland Research Laboratory St Lucia, Queensland, Australia E-mail: [email protected] Ramakoti Sadananda Rangsit University, Bangkok, Thailand E-mail: [email protected]
PRIMA has emerged as a major platform for academic and research exchange on agent technologies. The PRIMA workshop series was initiated as a workshop of the Pacific Rim International Conference in Artificial Intelligence (PRICAI) to provide a forum that would bring together research in the areas of agent technology and multi-agent systems, both in the Pacific Rim region and beyond. The inaugural workshop in the series was held in Singapore in 1998, with subsequent meetings in Kyoto (1999), Melbourne (2000), Taipei (2001), Tokyo (2002), Seoul (2003), Auckland (2004), Kuala Lumpur (2005) and Guilin (2006). At the 10th PRIMA in Bangkok in November 2007, the Steering Committee agreed that the series had grown in size and achieved a level of maturity to become a conference series of its own. It was therefore agreed that from Bangkok in 2007 PRIMA would stand for the Pacific Rim International Conference on Multi-Agent Systems. PRIMA 2007 received 102 valid submissions. Each submission was peerreviewed by at least three referees selected from the Program Committee. As a result of the selection process, 22 submissions were accepted as full research papers, yielding an acceptance rate of 22.22%. In addition the program included 11 application papers and 16 short papers. A special session on Agent-Oriented Software Engineering (AOSE) was organized by Graham Low from the University of New South Wales (Australia) and Ghassan Beydoun from the University of Wollongong (Australia), where papers were invited from the AOSE community, but put through the same rigourous reviewing process. The papers finally selected for PRIMA 2007 represent a range of themes – expanding the horizons of the agent metaphor, consistent with the way the computing landscape has moved from what was once stand-alone systems to the situation where the real power is realized through distributed, open and dynamic systems. We are faced with new opportunities and great challenges in a computing milieu where intelligent and autonomous agents interacting with one another solve problems of ever-growing complexity and difficulty. We congratulate and thank the authors of all the papers submitted and the reviewers who carefully reviewed them. We appreciate the academic institutions, especially Rangsit University of Thailand, for hosting and supporting this event. We acknowledge the support rendered by the sponsors, and the members of the committees for doing a splendid job.
November 2007
Aditya Ghose Guido Governatori Ramakoti Sadananda
Opening Address
It gives me great pleasure and honor to welcome you all to Rangsit University and to Thailand on the occasion of the 10th Pacific Rim International Workshop on Multi-Agent and Autonomous Agents. As I notice there is a strong distinguished academic community providing a sustained backing to the PRIMA workshop series. I am pleased to have served as Honorary Chairman. In that capacity and as President of Rangsit University, I extend a warm welcome to you all. Some of you here are also participating in the International Conference on IT for Empowerment (ITE2007). I extend my welcome to you too. I am a layman as far as the multi-agent systems are concerned. Like every one else I know only one kind of agent: the travel agent. He is the agent of the airlines to me. He enjoys considerable autonomy in the way he charges fees or the way he deals with the airlines. The travel agent is now increasingly embodied in the pages of the worldwide web, thus augmenting effectiveness and efficiency. Well, when you generalize and virtualize this mundane concept of “agency” it emerges as a mighty concept, transcending boundaries of all kinds. From the program I see your papers addressing a grand spectrum of areas – from human science to biology and from engineering to game theory. You are designing software agents and intelligent agents roaming around the Internet negotiating with one another. The Internet is now the media for the planet. We all operate on the Internet. I do not want to dwell on technical and philosophical aspects of your research. But let me talk to you about a common concern for all of us – no matter where you come from - the threat of climate change. Currently there are world-wide discussions on global warming and the possible serious consequences, means to prevent or mitigate it and/or maybe to face it. The recent Nobel awards recognize that the threat is real and largely manmade. There is an urgent need for collective action. We are in Bangkok and are among the most vulnerable—especially the place we are now assembled in. I ask you – every one of you – as researchers in information technologies, what you can offer. The situation facing us is complex – especially because while facing the climate crisis, we have to carry the unfinished responsibilities of development. There is still a huge segment of the world population that is undernourished, under-informed and living under unacceptable standards. There is a notion, especially following the Second World War, that typically innovations emerge from the war efforts. We are heading towards a war. But this is war of a different kind. It is not a war of conflict but one of cooperation. The global crisis with all its threats and horrors is pushing us to work together with greater interactions than ever before. You may design agents – the agents who represent us individually and collectively—with skills to negotiate win – win deals for all of us to face the impending manmade catastrophy. This is not utopian. There is boundless
VIII
Opening Address
potential in intelligent information systems when backed by the strength of the collective minds of all of us. I see great promise in agent research. I am pleased you have decided to assemble here and I welcome you to our university – especially to our main campus – a few kilometers north of here in the fertile land of Rangsit. Given your fertile mind, interaction with our young students and colleagues will germinate great ideas. I am happy to see that despite being busy in your path-breaking research, you are interested in student research. As an educationist I join you in working for students – in working for the future. I am delighted to see the doctoral mentoring program. This event will go a long way. One of the concerns of our research students, and even the faculty who advise them, is that being in small groups they face frustrating isolation. This is a bane in developing countries, including Thailand where it is hard to get peer-to-peer interaction. A meeting with distinguished researchers like you means a lot to them. It means a lot in terms of building their confidence and in refining their ideas. I am sure there is always benefit in talking to the innocent students whose innocent questions may open new agendas for research. Let me borrow from Shakespeare “it is twice blest: It blessth him that gives and him that takes”. Again with great pleasure and honor I welcome you all to my university, to the city of Angeles and to the land of smiles.
November 2007
Arthit Ourairat
Invited Speakers
Yaser S. Abu-Mostafa, CalTech, USA: Ensemble Learning One of the core ideas of multi-agents is the accomplishment of a collective task through the contributions of multiple units. The same idea arises in the context of computational learning, where a learning task is accomplished by contributions from an ensemble of learning units. Ensemble learning is a very active area of research, and some of its algorithms such as AdaBoost are in wide use in practical applications. This talk gives the background of ensemble learning and describes its main algorithms. Some of the theoretical questions in ensemble learning parallel those in multi-agents, such as the level of supervision and independence of the individual learners. The key question, however, is why ensemble learning works so well. Several attempts at answering the question theoretically have consequently been shown to be false. Many of these ideas are discussed and it is shown that it is not only a curiosity to answer this question, but an important step in the optimal design of ensemble learning algorithms.
Hideyuki Nakashima, Future University, Hakodate, Japan: Future of Multi-agent Systems Research with Eastern Philosophy Leon Sterling, University of Melbourne, Australia: Agent-Oriented Modelling for Complex Socio-technical Systems Agent-oriented software engineering has emerged as an active research area in the past decade. This talk argues that agent concepts are useful from the earliest stages of system conceptualization. The agent paradigm is useful even if systems are implemented without using agent concepts. Examples of agent modes are presented.
Conference Organization
Honorary Chair Arthit Ourairat
General Chair R. Sadananda
Program Chairs Aditya Ghose Guido Governatori
Doctoral Mentoring Chair The Duy Bui
Program Committee Yaser S. Abu-Mostafa Mike Barley Dencho Batanov Ghassan Beydoun Richard Booth Rafael H. Bordini Stephane Bressan The Duy Bui Suresh Chande Joongmin Choi John Debenham Joerg Denzinger Frank Dignum Virginia Dignum Alexis Drogoul Atilla Elci Paolo Giorgini Tuong Vinh Ho Michael C. Horsch Jane Hsu
Toru Ishida Kamalakar Karlapalem Minkoo Kim Yasuhiko Kitamura Jean-Luc Koning Ryszard Kowalczyk Aneesh Krishna Jaeho Lee Ho-fung Leung Chao-Lin Liu Lin Liu Jyi-Shane Liu Rey-Long Liu Alan Liu Graham Low Dickson Lukose Thomas Meyer S. B. Nair Hideyuki Nakashima Mehmet Orgun
Sascha Ossowski Vineet Padmanabhan Mikhail Prokopenko Maryam Purvis Antonino Rotolo R. Sadananda Abdul Sattar Hajime Sawmura Ali Selamat Liz Sonenberg Leon Sterling Toshiharu Sugawara Soo Von-Wun Michael Winikoff Wayne Wobcke Jung-Jin Yang Zili Zhang Dongmo Zhang Hans van Ditmarsch Leon van der Torre
Local Organization Chattrakul Sombattheera Chom Kimpan Damrong Indharameesup Esa Kujansuu Gyanam Sadananda Henna H¨alinen HwaJoon Kim Jaakko Lautarinne Jeroen Schedler Julia Jokinen Malivan Praditeera
External Reviewers Mohan Baruwal Chhetri Aliaksandr Birukou Jeff Blee Jan Broersen Volha Bryl Patrice Caire Chi-Kong Chan Timothy Cleaver Alan Colman Fabiano Dalpiaz Hind Fadil Taro Fukumoto Masabumi Furuhata Aurona Gerber Steve Goschnick Octavio Gutierrez Ramon Hermoso Traba Hoyoung Jeung Thomas Juan Kazuki Kobayashi George Koliadis Shusuke Kuribara Ka-man Lam
Jaeho Lee Louise Leenen Li Li Qiang Liu Lin Liu Qing Liu Graham Low Huiye Ma Wataru Makiguchi Melanie Middlemiss Juliana Mitchell-Wong Tomokazu Narita Philippe Pasquier Laurent Perrussel Bastin Tony Roy Savarimuthu Yohsuke Takahashi Matteo Vasirani Bao Quoc Vo Mengxiao Wu Zhang Yang Jung-Jin Yang
Table of Contents
Existence of Risk Strategy Equilibrium in Games Having No Pure Strategy Nash Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ka-man Lam and Ho-Fung Leung
MAGEFRAME: A Modular Agent Framework to Support Various Communication Schemas Based on a Self-embedding Algorithm . . . . . . . . Quintin J. Balsdon and Elizabeth M. Ehlers
25
Using Multiagent System to Build Structural Earth Model . . . . . . . . . . . . Beiting Zhu, Zahia Guessoum, Michel Perrin, Bertrand Braunschweig, Pierre Fery-Forgues, and Jean-Fran¸cois Rainaud
37
Agent-Supported Protein Structure Similarity Searching . . . . . . . . . . . . . . Dariusz Mrozek, Bo˙zena Malysiak, and Wojciech Augustyn
49
Merging Roles in Coordination and in Agent Deliberation . . . . . . . . . . . . . Guido Boella, Valerio Genovese, Roberto Grenna, and Leendert van der Torre
62
Planning Actions with Social Consequences . . . . . . . . . . . . . . . . . . . . . . . . . Hsueh-Min Chang and Von-Wun Soo
Coalition Structure Generation in Task-Based Settings Based on Cardinality Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . She-Xiong Su, Shan-Li Hu, Sheng-Fu Zheng, Chao-Feng Lin, and Xian-Wei Lai A Specialised Architecture for Embedding Trust Evaluation Capabilities in Intelligent Mobile Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Justin R. Pike, Elizabeth M. Ehlers, and Ockmer L. Oosthuizen
398
404
XVI
Table of Contents
Reasoning with Levels of Modalities in BDI Logic . . . . . . . . . . . . . . . . . . . . Jeff Blee, David Billington, and Abdul Sattar
410
A Distributed Computational Model for Mobile Agents . . . . . . . . . . . . . . . Med Amine Haddar, Ahmed Hadj Kacem, Yves M´etivier, Mohamed Mosbah, and Mohamed Jmaiel
Existence of Risk Strategy Equilibrium in Games Having No Pure Strategy Nash Equilibrium Ka-man Lam and Ho-Fung Leung Department of Computer Science and Engineering The Chinese University of Hong Kong {kmlam,lhf}@cse.cuhk.edu.hk
Abstract. Two key properties defining an intelligent agent are reactive and pro-active. Before designing an intelligent agent for any multi-agent system, we need to first understand how agents should behave and interact in that particular application, which can be done by modelling the application as a game. To analyze these games and to understand how decision-makers interact, we can use a collection of analytical tools known as Game Theory. Risk strategies is a new kind of game-theoretic strategy. Simulations in previous work have shown that agents using risk strategies are reactive as well as pro-active and thus have better performance than agents using other models or strategies in various applications. However, research on risk strategies has been focusing on formalization, application, and games having pure strategy Nash equilibrium. In this paper, we analyze a game having no pure strategy Nash equilibrium. We find that risk strategy equilibrium may exist even the game does not have pure strategy Nash equilibrium. We then summarize general conditions for the existence of risk strategy equilibrium. Simulation shows that agents using risk strategies also have better performance than agents using other existing strategies in a game having no pure strategy Nash equilibrium.
1
Introduction
Autonomous agents, also known as intelligent agents, are agents with reactivity and pro-activeness. Reactivity is the ability that an intelligent agent can react to the changes in the environment instantly, so as to achieve the design objectives. However, if an agent can just react to the environment, the agent will be directed by the environment and deviate from its design objectives very easily. So, an intelligent agent should be able to recognize opportunities and take the initiative to meet its goals. This is Pro-activeness. In order to design intelligent agents for a multi-agent system that are reactive and pro-active, we need to understand how agents should behave in the system environment. One way to do so is to model agent interaction in that environment as a game. Many human interactions and economic activities can be modeled as games, like auctions, negotiations, bidding, bargaining, and political competition. Game Theory is a tool we can use to model agent interaction as a game, analyze and understand how decision-makers interact in the game. A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 1–12, 2009. c Springer-Verlag Berlin Heidelberg 2009
2
K. Lam and H. Leung
In game theory, decision-making sometimes takes into account utility only, while sometimes also takes into account probability and uses expected utility maximization. However, experiments [6] show that most of the decision makers violate the Expected Utility Theory [3] when there is risk in the choices. This is because decision makers have attitudes toward risk. However, risk attitude is rarely considered in game analysis. Another key factor affecting a person’s decision-making is other people’s reputations [19]. At the same time, there are also various models of reputation in the context of multi-agent systems [5,15,17]. These models show that it is also good for agents to make decisions based on other agents’ reputations. Again, decision-making based on reputations is also rarely considered in game analysis. In [9], Lam and Leung propose a new model of decision-making, called attractiveness maximization, which uses risk attitude, probability and utility in decision-making. The idea is that instead of calculating the product of probability and utility, some people are attracted by the large amount of gain even the probability of gain is very low, while some people are resisted by the low probability of gain even the amount of gain is very large. Lam and Leung use the risk attitude of a person as a weight on the probability of gain and the amount of gain to calculate the attractiveness of an outcome and the choice in which the outcome has the maximum attractiveness is chosen. They show that the model of attractiveness maximization can be used to explain the Allais paradox [1], which violates the Expected Utility Theory. They also find some properties of the model of attractiveness maximization, which match the human behaviors. Then, Lam and Leung extend the model of attractiveness maximization and apply the extended model to infinitely repeated games [8,10]. In game decisionmakings, reputations of players are considered instead of probability. Then they propose a new kind of strategies and equilibrium called risk strategies and risk strategy equilibrium. The authors differentiate these new concepts with existing concepts in game theory, like pure strategies, mixed strategies, behavioral strategies, and trigger strategies. Simulation also shows that players using risk strategies outperform players using other game theoretic strategies. Add to this, Lam and Leung apply risk strategies to design intelligent agents for various multi-agent systems [7,12,13,14], including minority games, resource allocation and the famous Iterated Prisoner’s Dilemma. Simulations show that agents using risk strategies are able to react to the opponent as well as to proactively seek for payoff, which match the design objectives of intelligent agents. Simulations also show that agents using risk strategies have better performance than previous models or strategies. Through analyzing the Iterated Prisoner’s Dilemma [14], Lam and Leung also summarize several properties of risk strategy equilibrium in infinitely repeated games. However, those properties hold only for infinitely repeated games in which the constituent strategic game has pure strategy Nash equilibrium. Yet, not every strategic game has pure strategy Nash equilibrium. In this paper, we find out
Existence of Risk Strategy Equilibrium in Games Having No Pure Strategy
3
properties of risk strategy equilibrium in infinitely repeated games in which the constituent strategic game has no pure strategy Nash equilibrium. In the next section, we describe a background of risk strategies, risk strategy equilibrium and their properties. In section 3, analyze a motivating game and summarize the properties of risk strategy equilibrium in infinitely repeated games in which the constituent strategic game has no pure strategy Nash equilibrium. In section 4, we perform simulations and show that risk strategies work better than existing strategies. Lastly, We conclude the paper.
2
Background
2.1
Risk Strategies and Risk Strategy Equilibrium
As suggested by the Expected Utility Theory [3], the risk attitude of a decisionmaker affects how the decision-maker values the outcome of a choice. On the other hand, as suggested by the Prospect Theory [6], the risk attitude of a person affects how the person values the probability that an outcome would occur. However, risk attitude is rarely considered in game analysis. Another factor affecting a person’s decision-making is other people’s reputations [19], which are also rarely considered in game analysis. In [14], Lam and Leung give formal definitions of risk attitude and reputation and incorporate these two concepts into infinitely repeated games. A strategic game is a game in which the players choose their actions once and for all and the players choose their actions simultaneously. When a strategic game is repeated infinitely, the game is called an infinitely repeated game. Definition 1. An infinitely repeated game consists of: – N = {1, 2, . . . , n} is the set of players. – Ai is the set of available actions for player i. – Ui is the utility function for player i, where Ui (a1 , . . . , an ) returns the payoff of player i for the action profile (a1 , . . . , an ). – i is the preference relation for player i.1 The k th time that the constituent strategic game is repeated is called round k of the infinitely repeated game. In round k, player i chooses action aki and gets a utility of uki , where uki = Ui (ak1 , . . . , akn ). A sequence of action profiles (ak1 , . . . , akn )∞ k=1 forms an outcome O of an infinitely repeated game, in which player i gets a sequence of payoffs (uki )∞ k=1 . In [14], each player is associated with a risk attitude state at each round of the game.
1
There are different ways to define preference relations i [16]. In this paper, we use limit of means: an outcome O1 , in which player i gets a payoff sequence (pki )∞ k=1 , is player i gets a payoff sequence preferred by player i to another outcome O2 , in which T k k (qik )∞ k=1 , denoted as O1 i O2 , if and only if limT →∞ k=1 (pi − qi )/T > 0.
4
K. Lam and H. Leung
Definition 2. In an infinitely repeated game, the risk attitude state of a k player i at round k of the game, rski , is a 3-tuple < ri,p , Ri , Rj >, where: k – ri,p is the attitude of player i towards the payoff at round k. k k – Ri is a tuple < ri,i,a , . . . , ri,i,a > where ai ∈ Ai , which are the attitudes of 1 n player i towards his own reputations of choosing each of his available actions at round k. k k , . . . , ri,j,a > where aj ∈ Aj , j ∈ N and i = j, which – Rj is a tuple < ri,j,a 1 n are the attitudes of player i towards player j’s reputations of choosing each of player j’s available actions at round k.
each of the attitude is a real number in [0, 1]. This is to model the kind of human behavior shown in the experiments conducted by Kahneman and Tversky [6]. In these experiments, some people consider payoff as a more important factor in the decision-making process. They are said to be k k = 1, ri,i,a = 0 for utility-seeking. This can be represented by a player with ri,p i k all ai ∈ Ai and ri,j,aj = 0 for all aj ∈ Aj . In contrast, some people want to reduce risk as much as possible. They are said to be risk-averse. In a game, there are two ways that a player can reduce risk by considering reputations. First is to consider other players’ reputations. Usually, the higher the reputation of a player in choosing a certain action, the more likely the player will choose the same action in future encounter. So by considering other players’ reputations, a player can choose a suitable action in response, which reduces the risk undertake. k k = 0, ri,i,a = 0 for all ai ∈ Ai and This can be represented by a player with ri,p i k ri,j,aj = 1 for some aj ∈ Aj . Another way is to maintain one’s own reputation in choosing a certain action. In this way, a player can attract other players to choose some cooperative actions in future encounters so that potential risk can k k = 0, ri,i,a = 1 for be lowered. This can be represented by a player with ri,p i k some ai ∈ Ai and ri,j,a = 0 for all a ∈ A . j j j In [14], each available action of each player is associated with a reputation. Definition 3. In an infinitely repeated game, the reputation of a player i in at the end of round k of the game, choosing an action a ∈ Ai , obtained repki,a , is a real number in [0, 1], where a∈Ai repki,a = 1. This is given by the ratio of the number of times that the player chooses action a out of the previous k rounds of the game. Elements of a person’s risk attitude state are used as a weight of payoff and reputations to calculate the attractiveness of an action profile. An action leading to a higher payoff than other actions is more attractive to a utility-seeking player while an action responding to other players’ reputable actions or an action that can maintain ones own reputations is more attractive to a risk-averse player. The payoff matrix of a game is then transformed to an attractiveness matrix
Existence of Risk Strategy Equilibrium in Games Having No Pure Strategy
5
in every round k by Eqt. 1.2 Then, players choose the most attractive action in response to every other player’s most attractive action. k k k k Attki (ak1 , . . . , akn ) = ri,p uki + ri,i,a rep + ri,j,a repk−1 (1) k i,a i j j,ak i
j
j∈[1,n],j=i
According to [14], a risk strategy is a function assigning a risk attitude state to the player in each round of the game. When all players are using their best risk strategies in response to the other players’ best response risk strategies, the situation is called a risk strategy equilibrium. Definition 4. A risk strategy RSi , of a player i in an infinitely repeated game is a function that assigns a risk attitude state rski to the player in each round k. Definition 5. A risk strategy equilibrium of an infinitely repeated ∗ game is a profile RS ∗ of risk strategies with the property that O(RS−i , RSi∗ ) i ∗ ∗ O(RS−i , RSi ) for every risk strategy RSi of player i, where O(RS−i , RSi∗ ) is called the outcome of the risk strategy equilibrium. 2.2
Properties of Risk Strategy Equilibrium
In [14], Lam and Leung research some properties of risk strategy equilibrium for infinitely repeated games in which the constituent strategic game has pure strategy Nash equilibrium. Before we introduce the properties, we recall the definitions of Pareto optimum and domination from [14]. A repetition of the same action profile (a1 , . . . , an ) for t rounds is denoted as (a1 , . . . , an )t and a sequence of action profiles in which the players play (a11 , . . . , an1 ) for p1 rounds, . . . , (a1t , . . . , ant ) for pt rounds is denoted as ((a11 , . . . , an1 )p1 , . . . , (a1t , . . . , ant )pt ). Definition 6. A sequence of action profiles ((a11 , . . . , an1 )p1 , . . . , (a1t, . . . ,ant )pt) is a Pareto optimum if andonly if there exist no non-negative integers p1 ’, . . . , t t t pk = pk , such that pk Ui (a1k , . . . , ank ) ≥ pt ’, where k=1 k=1 k=1 t k=1 pk Ui (a1k , . . . , ank ) for all i ∈ N . p1 pt Definition 7. A sequence of action profiles ((a11 , . . . , a n1) , . . . ,(a1t, . . . ,ant) ) dominates an action profile (a∗1 , . . . , a∗n ) if and only if tk=1 pk Ui (a1k , . . . , ank ) t ≥ ( k=1 pk )Ui (a∗1 , . . . , a∗n ) for all i ∈ N .
Property 1. Let G =< N, (Ai ), (Ui ) > be a strategic game. An indefinite repetition of a sequence of pure strategy Nash equilibria is an outcome of a risk strategy equilibrium in the infinitely repeated game of G if and only if the sequence is a Pareto optimum. 2
As at the beginning of round k, player i does not know what action other players j (j ∈ [1, n], j = i) will choose in round k, player i considers other players’ reputations obtained at the end of round k − 1. rep0j,ak is the initial reputations, which repk−1 j,ak j
j
are given to each player at the beginning of the game. In contrast, player i considers his own reputations repki,ak that will be obtained at the end of round k if he chooses action aki in round k.
i
6
K. Lam and H. Leung Player 2
Player 1
Player 2
C
D
C
r, r
s, t
D
t, s
p, p
Player 1
Fig. 1. The Prisoner’s Dilemma
B
N
L
2, -2
-1, 0
T
1, 1
0, -1
Fig. 2. Lying Game
For example, in the Prisoner’s Dilemma as shown in Fig. 1 (with t = 5, r = 3, p = 1, s = 0), (D, D) is a pure strategy Nash equilibrium. Playing (D, D) in every round of the Iterated Prisoner’s Dilemma is also a risk strategy equilibrium. Property 2. Let G =< N, (Ai ), (Ui ) > be a strategic game. An indefinite repetition of a sequence of action profiles (but not pure strategy equilibria) is an outcome of a risk strategy equilibrium in the infinitely repeated game of G if and only if the sequence is a Pareto optimum, and dominates all the pure strategy Nash equilibria in G. In the Iterated Prisoner’s Dilemma, playing (C, C) in every round of the game is Pareto optimum. Also, as Ui (C, C) > Ui (D, D) for all i ∈ N , playing (C, C) in every round of the game dominates the pure strategy Nash equilibrium (D, D). This Pareto optimum is an outcome of a risk strategy equilibrium. Also, in a variation of the Iterated Prisoner’s Dilemma (with t = 50, r = 3, p = 1, s = 0), playing (C, D) and (D, C) alternatively is also a Pareto optimum. As Ui (C, D) + Ui (D, C) > 2Ui (D, D) for all i ∈ N , playing (C, D) and (D, C) alternatively dominates the pure strategy Nash equilibrium (D, D). This Pareto optimum is also an outcome of a risk strategy equilibrium.
3 3.1
Risk Strategy Equilibrium in Games Having No Pure Strategy Nash Equilibrium Motivating Example: Lying Game
In [11], Lam and Leung propose a trust/honesty model for the semi-competitive environment. In such an environment, agents have incentives to tell the truth as well as to tell lies. However, their discussions are only on a simulative approach. In this paper, we model the semi-competitive environment as a Lying game and give game-theoretic analysis. In the Lying Game, player 1 can choose to lie or to tell the truth, denote as L and T respectively, while player 2 can choose to believe or not to believe player 1, denote as B and N respectively. The preference relations over the action profiles can be expressed as payoffs, as shown in Fig. 2. If player 1 chooses L, it is better for player 2 to choose N . If player 2 chooses N , it is better for player 1 to choose T . If player 1 chooses T , it is better for player 2 to choose B. If player 2 chooses B, it is better for player 1 to choose L. So, there is no pure strategy Nash equilibrium in this game.
Existence of Risk Strategy Equilibrium in Games Having No Pure Strategy
Player 2 B Player 1
L T
1, 1+rep
Player 2 N
0, -2
0, 0 0
1,T
7
1, -1+rep
Player 1 0
1,T
Fig. 3. First round of the Lying Game, transformed using Eqt. 1 and Eqt. 2
B
N
L
0, -2
0, 0
T
1, 2
1, 0
Fig. 4. Second round of the Lying Game, transformed using Eqt. 1 and Eqt. 2
k k k k k Let us denote rski =< ri,p , < ri,1,L , ri,1,T >, < ri,2,B , ri,2,N >> as player i’s risk attitude state in round k, where i = 1, 2. If player 1 wants to be a good man (not to lie) and player 2 chooses his action based on whether player 1 is a good man. Both players care player 1’s reputation in choosing T . However, if player 1 lies in the previous round, then player 2 will be cautious to player 1 and care about player 1’s reputation in choosing L and his own reputation in choosing N . This is represented by the following risk strategies:
for all k for k = 1 =L for k ≥ 2, and ak−1 1 otherwise
(2)
The transformed game of the first round will be the one as shown in Fig. 3. It is more attractive for player 1 to choose T no matter what player 2 chooses. If player 1 chooses T , it is more attractive for player 2 to choose B as 1 + rep01,T > −1 + rep01,T . So, player 1 chooses T and player 2 chooses B in round 1. At the end of round 1, rep11,T = rep12,B = 1 and rep11,L = rep12,N = 0. Both players do not change their risk attitudes according to Eqt. 2. Then, the transformed game of the second round will be the one as shown in Fig. 4. So in round 2, it is also the most attractive for player 1 to choose T no matter what player 2 chooses. If player 1 chooses T , it is also more attractive for player 2 to choose B as 2 > 0. Again, both players do not change their risk attitudes according to Eqt. 2. Also, rep21,T = rep22,B = 1 and rep21,L = rep22,N = 0 at the end of round 2. Then, the decision-making remains the same in round k > 2. As a result, player 1 chooses T and player 2 chooses B in every round of the game. In this case, both players get a payoff of k in k rounds of the game. Since player 2, using the risk strategy as shown by Eqt. 2, effectively chooses B in every round of the game, then player 1 can increase his one-round payoff by using other risk strategies such that he will choose L instead of T in any round of the game. Say player 1 uses other risk strategies such that he chooses L in round h. Then, player 2 changes his risk attitudes in round h + 1 according to Eqt. 2. Then it will be more attractive for player 2 to choose N starting from round h + 1 no matter what player 1 chooses. Then the best response for player 1 is to choose T starting from round h + 1. In this case, player 1 gets a payoff of 1(h − 1) + 2 = h + 1 in k rounds of the game, where k > h. As limk→∞ (h+1−k)/k < 0, player 1 cannot increase his payoffs for the whole game
8
K. Lam and H. Leung
by using other risk strategies. On the other hand, player 2 is already getting the highest payoff by using Eqt. 2. This is a risk strategy equilibrium. 3.2
Existence of Risk Strategy Equilibrium in Games Having No Pure Strategy Nash Equilibrium
From the Lying Game, we find that a risk strategy equilibrium may exist in an infinitely repeated game even the constituent strategic game has no pure strategy Nash equilibrium. Property 3. Let G =< N, (Ai ), (Ui ) > be a strategic game with no pure strategy Nash equilibrium. An indefinite repetition of a sequence of action profiles is an outcome of a risk strategy equilibrium in the infinitely repeated game of G if and only if the sequence is a Pareto optimum, and dominates all other non Pareto optimums in G. Outline of the proof: Let G =< N, (Ai ), (Ui ) > be a strategic game. Let a sequence of action profiles be ((a11 , . . . , an1 )p1 , . . . , (a1t , . . . , ant )pt ). A risk strategy equilibrium is: ⎧ k k k k k < ri,p = 0, < ri,i,a = 1, ri,i,a = 0 >, < ri,j,a = 1, ri,j,a = 0 >> ⎪ it i jt j ⎪ ⎪ ⎪ + h(p + · · · + p ) + 1 for k = p t−1 1 t ⎨ k k k k k = 0, < ri,i,a = 0, ri,i,a = 1 >, < ri,j,a = 0, ri,j,a = 1 >> (3) rski = < ri,p it i jt j ⎪ k−1 k−1 ⎪ ⎪ for ai = ait and aj = ajt ⎪ ⎩ k−1 rsi otherwise for all i, j ∈ N , i = j, ai ∈ Ai and ai = ait , aj ∈ Aj and aj = ajt , where h is a non-negative integer and p0 = 0. Then, the players choose (a11 , . . . , an1 ) for p1 rounds, . . . , (a1t , . . . , ant ) for pt rounds repeatedly. Suppose other players use Eqt. 3 and choose action ajt but player i uses another risk strategy such that he can increase his payoff by action ai instead of ait in any one round of the game. If so, other players will change their risk attitudes according to Eqt. 3 and will not choose ajt in succeeding rounds of the game. As the sequence of action profiles ((a11 , . . . , an1 )p1 , . . . , (a1t , . . . , ant )pt ) is a Pareto optimum, player i cannot increase his payoff by other risk strategies. As a result, Eqt. 3 is a risk strategy equilibrium. The sequence of action profiles ((a11 , . . . , an1 )p1 , . . . , (a1t , . . . , ant )pt ) must be a Pareto optimum. Otherwise, exists some non-negative integers p1 ’, . . . , t t there t p = p , such that p U (a pt ’, where k i 1k , . . . , ank ) ≥ k=1 k=1 k k=1 k t p U (a , . . . , a ) for all i ∈ N . Then the players can increase their paynk k=1 k i 1k offs by using a risk strategy similar to Eqt. 3 by replacing p1 with p1 , . . . , pt with pt . Suppose the sequence of action profiles ((a11 , . . . , an1 )p1 , . . . , (a1t , . . . , ant )pt ) does not dominate one of the non Pareto optimum (a1 , . . . , an ), we have t t k=1 pk Ui (a1k , . . . , ank ) ≥ ( k=1 pk )Ui (a1 , . . . , an ) for some i ∈ N . Then player i can increase his payoff by using a risk strategy which considers only his own reputation in choosing action ai and other player j’s reputation in choosing action aj .
Existence of Risk Strategy Equilibrium in Games Having No Pure Strategy
9
Property 4. Let G =< N, (Ai ), (Ui ) > be a strategic game with no pure strategy Nash equilibrium. If all the outcomes in G are Pareto optimums, then an indefinite repetition of a sequence of action profiles is an outcome of a risk strategy equilibrium in the infinitely repeated game of G if and only if the sequence is a Pareto optimum. By Property 3, an indefinite repetition of a sequence of action profiles is an outcome of a risk strategy equilibrium in the infinitely repeated game of G if and only if the sequence is a Pareto optimum and dominates all other non Pareto optimums in G. As there is no non Pareto optimum in G, an indefinite repetition of a sequence of action profiles is an outcome of a risk strategy equilibrium in the infinitely repeated game of G if and only if the sequence is a Pareto optimum.
4
Simulation
In the last section, we find that a game is in equilibrium when both players use risk strategies, even the game has no pure strategy Nash equilibrium. However, the idea of equilibrium is that players are playing the best response given the other players’ strategies, which is not ideal in practice. In real situations, players usually do not know in advance what strategies that other players are playing. In this section, we will show that not only a game is in equilibrium when players use risk strategies, players also get high payoffs by using risk strategies without knowing in advance what strategies that other players are playing. Simulation is done to compare the performance of various strategies. We have 6 strategies for player 1 (and for player 2): 1. 2. 3. 4.
Random: chooses L(B) or T (N ) with probability 0.5. All L(All B): chooses only L(B). All T(All N): chooses only T (N ). Trigger: initially chooses T (B), but chooses L(N ) whenever the other player chooses N (L) in the previous round, then resume to T (B). 5. TFT: initially chooses T (B), and continue to choose T (B) unless the other player chooses N (L) in the previous round. 6. Risk: uses Eqt. 4(Eqt. 5). ⎧ < 0, < 0, 1 >, < 0, 0 >> for k = 1 ⎪ ⎪ ⎨ < 1, < 0, 0 >, < 0, 0 >> for k ≥ 1, rsk1 = < 0, < 0, 1 >, < 1, 0 >> for k ≥ 1, ⎪ ⎪ ⎩ k−1 rs1 otherwise ⎧ < 0, < 0, 0 >, < 1, 0 >> for k = 1 ⎪ ⎪ ⎨ < 1, < 0, 0 >, < 0, 0 >> for k ≥ 1, rsk2 = < 0, < 1, 0 >, < 0, 1 >> for k ≥ 1, ⎪ ⎪ ⎩ k−1 rs1 otherwise
repk−1 2,B = 1 k−1 repk−1 2,B ≥ rep2,N
(4)
repk−1 1,L = 1 k−1 repk−1 1,L ≥ rep1,T
(5)
The idea of Eqt. 4 and Eqt. 5 is to be cooperative and care the player’s own reputation in choosing T (or B) initially. Then becomes utility-seeking if player
2’s (or player 1’s) reputation in choosing B (or L) is equal to 1 so as to proactively seek for payoff. If player 2 (or player 1) does not always choose B (or L), but player 2’s (or player 1’s) reputation in choosing B (or L) is greater than that of choosing N (or T ), then this risk strategy will care the player’s own reputation in choosing T (or N ) and player 2’s (or player 1’s) reputation in choosing B (or L) so as to seek for cooperation (or reduce loss). It can be verified that the game is in risk strategy equilibrium if both players use Eqt. 4 and Eqt. 5. Each strategy of player 1 will encounter each strategy of player 2 for a game. Each game has 1000 rounds. The result is shown in Fig. 5. Each tuple in the figure denotes the payoffs of player 1 and player 2 in a game respectively. The sum of player 1’s payoffs and the sum of player 2’s payoffs of each strategy are shown at the end of each row and column respectively. From the figure, we can see that player 1 gets the highest payoff by using risk strategy, which is about 25% higher than the second largest payoff obtained by using All T strategy. Player 2 also gets the highest payoff by using risk strategy, though it is closely to the second and the third highest payoff. In all games, players using risk strategies are able to get the highest payoff or a payoff very close to the highest payoff. In particular, if player 2 uses All B strategy, the best response is to use All L strategy, getting a payoff of 2000. We can see that player 1 using risk strategy gets a payoff of 1999, which is very close to the best response. This is because reputations of the opponent are taken into account in decision-making. So if the opponent always chooses L, the player using risk strategy will become payoffseeking so as to pro-actively seek for payoff. In contrast, other strategies cannot seek for payoff pro-actively. However, if the opponent is using TFT or Trigger strategies, the opponent will response with an N . As the risk strategy is reactive, the player using risk strategy soon become cooperative again. If we know in advance what strategy the opponent will use, we can always choose the best strategy to response. For example, if we know that player 2 is using All B, then player 1 should use All L, but it may not be the best response to other strategies. In practice, players do not know what strategies the other
Existence of Risk Strategy Equilibrium in Games Having No Pure Strategy
11
players use. However, we can see that players using risk strategies are able to get the highest payoff or a payoff very close to the highest payoff in all cases, without knowing what strategies the opponent is using. This is because we take into account reputations as well as payoff in making decisions. Also, risk strategies are reactive and pro-active, meeting the design objectives of an intelligent agent. We use TFT and Trigger strategies for comparison because TFT has been a very good strategy for a long time [4] and Trigger strategy is a famous strategy for repeated games [2,18]. However, TFT and Trigger have good performance only if the other player also uses the same strategy. The performance may not be good if the other player uses other strategies. For example, we can see from Fig. 5 that if player 1 gets a payoff of −999 if he uses TFT or Trigger while player 2 uses All N. Another disadvantage of Trigger strategy is that players need to make prior agreements before the game start, which is impractical. On the other hand, players using risk strategy get high payoff no matter what strategies the other player uses. Also, piror agreements are not needed.
5
Conclusion
To design intelligent agents that are reactive and pro-active, we can use a collection of analytical tools known as Game Theory. Risk strategies is a new kind of game-theoretic strategy, which is shown to have better performance than agents using other models or strategies in various applications. Properties of risk strategy equilibrium in games having pure strategy Nash equilibrium are researched in previous work. In this paper, we analyze the Lying game, which is a game having no pure strategy Nash equilibrium. In this kind of games, we find that risk strategy equilibrium may still exist and we summarize general conditions for the existence. Simulation show that agents using risk strategies also have better performance than agents using other existing strategies in the Lying game. This is not only because reputations as well as payoff are taken into account in decision-making. The more important, risk strategies meet the design objectives of an intelligent agent, being reactive and pro-active.
Acknowledgement The work described in this paper was supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. 413306).
References 1. Allais, M.: Le comportement de l’homme rationnel devant le risque: 367 critique des postulats et axiomes de l’ecole americaine. Econometrica 21, 503–546 (1953) 2. Aumann, R.J., Shapley, L.S.: Long-term competition - a game-theoretic analysis. In: Essays in Game Theory, pp. 1–15. Springer, New York (1994)
12
K. Lam and H. Leung
3. Bernoulli, D.: Exposition of a new theory on the measurement of risk. Econometrica 22(1), 23–36 (1954) 4. The Prisoner’s Dilemma Competition (2005), http://www.prisoners-dilemma.com/ 5. Glass, A., Grosz, B.: Socially conscious decision-making. In: Proceedings of the Fourth International Conference on Autonomous Agents, pp. 217–224 (2000) 6. Kahneman, D., Tversky, A.: Prospect theory: An analysis of decision under risk. Econometrica 47(2), 263–291 (1979) 7. Lam, K.M., Leung, H.F.: Behavioral predictors and adaptive strategy for minority games. Physica A (submitted) 8. Lam, K.M., Leung, H.F.: Risk strategies and risk strategy equilibrium in agent interactions modeled as normal repeated 2 by 2 risk games. In: The Eighth Pacific Rim International Workshop on Multi-Agents (2005) 9. Lam, K.M., Leung, H.F.: Expected utility maximization and attractiveness maximization. In: Shi, Z.-Z., Sadananda, R. (eds.) PRIMA 2006. LNCS, vol. 4088, pp. 638–643. Springer, Heidelberg (2006) 10. Lam, K.M., Leung, H.F.: Formalizing risk strategies and risk strategy equilibrium in agent interactions modeled as infinitely repeated games. In: Shi, Z.-Z., Sadananda, R. (eds.) PRIMA 2006. LNCS, vol. 4088, pp. 138–149. Springer, Heidelberg (2006) 11. Lam, K.M., Leung, H.F.: A trust/honesty model with adaptive strategy for multiagent semi-competitive environments. Autonomous Agents and Multi-Agent Systems 120(3), 293–359 (2006) 12. Lam, K.M., Leung, H.F.: An adaptive strategy for minority games. In: The 6th International Joint Conference on Autonomous Agents and Multiagent Systems (2007) 13. Lam, K.M., Leung, H.F.: An adaptive strategy for resource allocation modeled as minority game. In: The First IEEE International Conference on Self-Adaptive and Self-Organizing Systems, pp. 193–202 (2007) 14. Lam, K.M., Leung, H.F.: Incorporating risk attitude and reputation into infinitely repeated games and an analysis on the iterated prisoner’s dilemma. In: The 19th IEEE International Conference on Tools with Artificial Intelligence (2007) 15. Mui, L., Mohtashemi, M., Halberstadt, A.: A computational model of trust and reputation. In: Proceedings of 35th Hawaii International Conference on System Science (2002) 16. Osborne, M.J., Rubinstein, A.: A Course in Game Theory. MIT Press, Cambridge (1994) 17. Rubiera, J.C., Lopez, J.M.M., Muro, J.D.: A fuzzy model of reputation in multiagent systems. In: Proceedings of the Fifth International Conference on Autonomous Agents, pp. 25–26 (2001) 18. Rubinstein, A.: Equilibrium in supergames. In: Essays in Game Theory, pp. 17–27. Springer, New York (1994) 19. Sartain, A.Q., North, A.J., Strange, J.R., Chapman, H.M.: Psychology – Understanding Human Behavior. McGraw-Hill Book Company, New York (1962)
Multiagent Planning with Trembling-Hand Perfect Equilibrium in Multiagent POMDPs Yuichi Yabu, Makoto Yokoo, and Atsushi Iwasaki Graduate School of ISEE, Kyushu University, 744 Motooka, Nishi-ku, Fukuoka, 819-0395, Japan [email protected], {yokoo,iwasaki}@is.kyushu-u.ac.jp
Abstract. Multiagent Partially Observable Markov Decision Processes are a popular model of multiagent systems with uncertainty. Since the computational cost for finding an optimal joint policy is prohibitive, a Joint Equilibrium-based Search for Policies with Nash Equilibrium (JESP-NE) is proposed that finds a locally optimal joint policy in which each policy is a best response to other policies; i.e., the joint policy is a Nash equilibrium. One limitation of JESP-NE is that the quality of the obtained joint policy depends on the predefined default policy. More specifically, when finding a best response, if some observation have zero probabilities, JESP-NE uses this default policy. If the default policy is quite bad, JESP-NE tends to converge to a sub-optimal joint policy. In this paper, we propose a method that finds a locally optimal joint policy based on a concept called Trembling-hand Perfect Equilibrium (TPE). In finding a TPE, we assume that an agent might make a mistake in selecting its action with small probability. Thus, an observation with zero probability in JESP-NE will have non-zero probability. We no longer use the default policy. As a result, JESP-TPE can converge to a better joint policy than the JESP-NE, which we confirm this fact by experimental evaluations. Keywords: Multiagent systems, Partially Observable Markov Decision Process, Nash equilibrium, Trembling-hand perfect equilibrium.
1
Introduction
Multiagent systems are increasingly being applied to such critical applications as disaster rescues, distributed unmanned air vehicles (UAV), and distributed sensor nets that demand robust, high-performance designs [1,2,3]. In these applications, we need to consider the uncertainty arising from various sources such as partial observability, imperfect sensing, etc. Multiagent Partially Observable Markov Decision Processes (Multiagent POMDPs) are emerging as a popular approach for modeling multiagent teamwork in the presence of uncertainty [4,5,6]. In a single-agent POMDP, a policy of an agent is a mapping of an agent’s observation history to actions. The goal is to find an optimal policy that gives the highest expected reward. In a Multiagent POMDP, the goal is to find an optimal A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 13–24, 2009. c Springer-Verlag Berlin Heidelberg 2009
14
Y. Yabu, M. Yokoo, and A. Iwasaki
joint policy of agents. Unfortunately, as shown by Bernstein et al. [7], the problem of finding an optimal joint policy for a multiagent POMDP is NEXP-Complete if no assumptions are made about the domain conditions. Therefore, a practical policy generation method requires to sacrifice optimality to some extent. Nair et al. [8] propose an algorithm called Joint Equilibrium-based Search for Policy with Nash Equilibrium (JESP-NE) that computes the locally optimal joint policy within a practical runtime. JESP-NE finds a locally optimal joint policy in which each policy is a best response to other policies, i.e., a Nash equilibrium. One limitation of JESP-NE is that the quality of the obtained joint policy depends on the predefined default policy. Assume that there are agent A and B. We first fix agent A’s policy and find the best response of agent B. There is a chance that some observation of agent B has zero probability (with the fixed A’s policy). In such a case, JESP-NE simply assigns default policy after observation with zero probability. This part of the policy does not affect the expected reward, since the probability that the part of the policy becomes active is zero. However, after finding the best response of agent B, we fix agent B’s policy and find the best response of agent A. In this case, part of B’s policy (where the default policy is assigned) can be active, since the policy of agent A will change. If the default policy is quite bad, the best response of agent A might be almost identical to the previous policy, so this part of B’s policy remains inactive. As a result, JESPNE converges to a sub-optimal joint policy, which is far from an optimal joint policy. In this paper, we propose a method that finds a locally optimal joint policy based on a concept called Trembling-hand Perfect Equilibrium (TPE) [9]. In finding a TPE, we assume that an agent might make a mistake in selecting its action with small probability. Thus an observation that has zero probability in JESP-NE will have non-zero probability. Therefore, we assign a best response policy rather than the default policy in this part. As a result, the JESP-TPE can converge to a better joint policy than JESP-NE. The experimental results show that JESP-TPE outperforms JESP-NE in settings where the default policy is not good enough. In the rest of this paper, we first show the multiagent POMDP model used in this paper (Section 2). Next we show an illustrative example, i.e., the multiagent tiger problem (Section 3). Then we describe JESP-NE (Section 4) and JESPTPE (Section 5). Finally, we show the comparison between the JESP-NE and JESP-TPE through experimental evaluations (Section 6).
2
Model
We follow the Markov Team Decision Problem (MTDP) model [10] as a concrete description of a multiagent POMDP. Given a team of n agents, an MTDP is defined as a tuple: S, A, P, Ω, O, R. S is a finite set of world states {s1 , s2 , . . . , sm }. A = Π1≤i≤n Ai , where A1 , . . . , An are the sets of actions for agents 1 to n. A joint action is represented as a1 , . . . , an . P (si , a1 , . . . , an , sf ), the transition
Multiagent Planning with Trembling-Hand Perfect Equilibrium
15
function, represents the probability that the current state is sf , if the previous state is si and the previous joint action is a1 , . . . , an . Ω = Π1≤i≤n Ωi is the set of joint observations where Ωi is the set of observations for agent i. O(s, a1 , . . . , an , ω), the observation function, represents the probability of joint observation ω ∈ Ω, if the current state is s and the agents’ previous joint action is a1 , . . . , an . We assume that an agent’s observations are independent of others’ observations. Thus the observation function can be expressed as: O(s, a1 , . . . , an , ω) = O1 (s, a1 , . . . , an , ω1 ) · . . . · On (s, a1 , . . . , an , ωn ). Each agent i forms a belief state, bti ∈ Bi , based on its observations seen through time t, where Bi represents the set of possible belief states for the agent. An agent relies on its state estimator function to update its belief state, given the latest observation. Finally, the agents receive a single, immediate joint reward R(s, a1 , . . . , an ) that is shared equally. This joint reward function is central to the notion of teamwork in a MTDP. Each agent i chooses its actions based on its local policy, πi , which is a mapping of its observation history to an action. Since we assume agents act as a team, i.e., they are cooperative, we don’t need to consider non-deterministic/ probabilistic policies. Using a Non-deterministic policy (in other words, using a mixed strategy) makes sense if there exists some adversarial agent. Thus, at time t, agent i will perform action πi (ω ti ) where ω ti = ωi1 , . . . , ωit refers to an observation history of agent i. The important thing to note is that in this model, execution is distributed but planning is centralized. Thus agents don’t know each other’s observations and actions at runtime but they know each other’s policies.
3
Example
We consider a multiagent version of the classic tiger problem introduced in [8]. Two agents are in front of the two rooms and its doors: “left” and “right”. Behind one door lies a hungry tiger and behind the other lies untold riches but the agents do not know the position of either. Thus, S = {SL, SR}, indicating behind which door the tiger is present. The agents can jointly or individually open either door. In addition, the agents can independently listen for the presence of the tiger. Thus, A1 = A2 = {OpenLef t, OpenRight, Listen}. The transition function P , specifies that every time either agent opens one of the doors, the state is reset to SL or SR with equal probability, regardless of the action of the other agent (see Table 1). However, if both agents choose Listen, the state is unchanged. The observation function O will return either HL or HR with different probabilities depending on the joint action taken and the resulting world state (see Table 2). For example, if both agents listen and the tiger is behind the left door (state is SL), each agent receives observation HL with probability 0.85 and HR with probability 0.15. Reward function R returns a joint reward (see Table 3). For example, the injury sustained if they opened the door to the tiger is less severe if they open that door jointly than if they open the door alone.
16
Y. Yabu, M. Yokoo, and A. Iwasaki
Table 1. Transition function P : * corresponds with either OpenLeft or OpenRight Action/Transition SL → SL SL → SR (SR → SR) (SR → SL) ∗, ∗ or Listen 0.5 0.5 ∗ or Listen, ∗ 0.5 0.5 Listen, Listen 1.0 0.0
Table 2. Observation function O: * of state corresponds with either SL or SR, * of action corresponds with OpenLeft or OpenRight State SL SR ∗ ∗
Action Listen, Listen Listen, Listen ∗, ∗ or Listen ∗ or Listen, ∗
HL 0.85 0.15 0.0 0.0
HR 0.15 0.85 0.0 0.0
Reset 0.0 0.0 1.0 1.0
Table 3. Reward function R: c corresponds with listen cost, d corresponds with tiger cost, and d/2 means the cost when agents jointly open door to tiger Action/State OpenRight, OpenRight OpenLef t, OpenLef t OpenRight, OpenLef t OpenLef t, OpenRight Listen, Listen Listen, OpenRight OpenRight, Listen Listen, OpenLef t OpenLef t, Listen
4
SL +20 −d/2 −d −d −c +10 +10 −d −d
SR −d/2 +20 −d −d −c −d −d +10 +10
JESP with Nash Equilibrium
In this section, we briefly explain Joint Equilibrium-based Search for Policies with Nash Equilibrium (JESP-NE) [8], which introduces the Nash equilibrium concept as a locally optimal approach. The Nash equilibrium is a solution concept of a game where no player has an incentive to unilaterally change strategy. The key idea in JESP-NE is finding the policy to maximize expected reward for other fixed agent’s policies. This process is continued until the joint policy becomes a Nash equilibrium. Algorithm 1 describes the JESP approach for n agents, given an initial belief state b and finite time horizon T . The key idea behind JESP is focusing on the decision problem for one agent at a time and keeping the policies of the other agent fixed; then the free agent faces a complex but normal single-agent POMDP. We repeat this process until we reach an equilibrium. Within each iteration of JESP, agent i calls the BestResponse function (line 5 in Algorithm 1) to find an individual policy that maximizes the expected joint reward given that the policies of the other agents are fixed. The problem of finding such a best response policy is equivalent to solving a single-agent POMDP. However, a belief state that stores distribution as in the single-agent case, is not a sufficient statistic because agent i must also reason about the other agents’ choice of actions, which, in turn, depend on their observation histories. Thus,
Multiagent Planning with Trembling-Hand Perfect Equilibrium
initialize V, π1 , . . . , πn repeat convergence ← true for i ← 1 to n do V , πi ← BESTRESPONSE(b, π¬i , T ) if πi = πi then convergence ← f alse V, πi ← V , πi until convergence return V, π1 , . . . , πn
Algorithm 2. BESTRESPONSE(b, πj , T ) 1: 2: 3: 4:
val ← GETVALUE(b, b, πj , 0, T ) initialize πi F IN DP OLICY (b, b, , πj , 0, T ) return val, π1
at each time t, agent i reasons about the tuple eti = st , Πj=i ωj t , where Πj=i ωjt represents the observation histories of the other agents. Then multiagent belief state B for agent i given the distribution over initial state b(s) is defined as: , b) Bit (eti ) = Bit (st , Πj=i ω tj ) = P r(st , Πj=i ω tj |ω ti , at−1 j Having fixed the policy of agent j (j = i), the best response for agent i can be computed using Algorithm 2. Following the model of the single agent value iteration algorithm, central to dynamic program is the value function over a T -step finite horizon. Value function Vt (B t , b) represents the expected reward that the team will receive, starting from the most recent belief state b and agent i following an optimal policy from the t-th step. Algorithm 3 constructs the optimal policy by maximizing the value function over possible action choices: Vt (Bit , b) = max Vtai (Bit , b) ai ∈Ai
(1)
Function Vtai can be computed using the GETVALUEACTION function as follows: a
Vt i (Bit , b) =
et
Bit (eti ) · (R(st , ai , πj (ωj , T − t)) +
P r(ωit+1 |Bit , ai ) · Vt+1 (Bit+1 )) (2)
t+1 ∈Ωi ω i
Bit+1 is the belief state updated after performing action ai and observing ωit+1 and is computed using the UPDATE function.
if t ≥ T then return 0 if Vt (B t , b) is already recorded then return Vt (B t , b) best ← −∞ for all a1 ∈ A do value ← GETVALUEACTION(B t , ai , b, πj , t, T ) Vtai (B t , b) ← value if value > best then best ← value Vt (B t , b) ← best return best
JESP with Trembling-Hand Perfect Equilibrium
In this section, we introduce Joint Equilibrium-based Search for Policies with Trembling-hand Perfect Equilibrium (JESP-TPE), which finds a locally optimal joint policy that corresponds to a Trembling-hand Perfect Equilibrium (TPE) [9]. 5.1
Nash Equilibrium and Trembling-Hand Perfect Equilibrium
JESP-NE obtains a joint policy that corresponds to a Nash equilibrium: i.e., each agent’s policy is a best response to other agents’ policies. In a Nash equilibrium, no agent has an incentive to unilaterally change strategy, assuming that no other agent changes its policies. A TPE 1 is proposed as a refinement of the concept of a Nash equilibrium. In a TPE, we assume each agent’s policy is a best response to other agents’ policies, even if other agents might deviate from the given policies with small probability . Definition 1 (Perturbed policy). A perturbed policy πit for a deterministic policy πi is a stochastic policy defined as follows: if πi chooses action ak in a certain situation, then, for the same situation, πit chooses ak with probability 1 − , and chooses each action aj (where j = k) with probability /(|Ai | − 1). Definition 2 (Trembling hand perfect equilibrium). For two agents, a combination of policies π1 , π2 is a TPE if the following conditions hold for all deterministic policies π1 , π2 , and π1t , π2t , where π1t , π2t are perturbed policies of π1 , π2 , respectively. JointReward(π1 , π2t ) ≥ JointReward(π1 , π2t )
(3)
JointReward(π1t , π2 )
(4)
JointReward(π1t , π2 ) 1
≥
The definition of a trembling hand perfect equilibrium used in the game theory literature is quite complicated and there exist several alternative definitions [11]. Here, we use a rather simplified definition that matches our model.
Multiagent Planning with Trembling-Hand Perfect Equilibrium
19
Table 4. Policy of agents i and j with finite horizon T = 2 t=0 t=1 observation action observation action (n/a) (Listen, Listen) HL (OpenRight, OpenRight) HR (OpenLef t, OpenLef t) Reset (OpenLef t, OpenRight)
As long as is small enough, if (3) and (4) are satisfied, the following conditions also hold. JointReward(π1 , π2 ) ≥ JointReward(π1 , π2 ) JointReward(π1 , π2 ) ≥ JointReward(π1 , π2 ) Thus, a TPE is also a Nash equilibrium, but not vice versa. First, let us describe a Nash equilibrium in the tiger problem. Table 4 shows a policy for agents i and j in a finite horizon, a listen cost, and a tiger cost (T = 2, c = 2, d = 40). This policy in Table 4 constitutes a Nash equilibrium, because each action of agent i is the best response against j. However, this policy does not satisfy a trembling-hand perfect equilibrium concept. Here, assume that agent j chooses OpenLeft by accident at the first step. Then both agents observe RESET. Thus agents i and j choose OpenLeft and OpenRight, respectively and receive rewards of −40: one agent encounters a tiger, while the other finds treasure. Therefore, this policy does not satisfy TPE when an agent may deviate from the given policy with small probability. Next, we consider a policy that constitutes a TPE. If we change the tuple of the actions for RESET as the observation with step t = 1 to (Listen, Listen) as the tuple of the actions, all agents receive a −2 reward, which is greater than the other tuple’s reward. Therefore, the changed policy constitutes a TPE. Notice that a TPE always constitutes a Nash equilibrium, but not vice versa. 5.2
JESP-TPE
This subsection describes the details of the proposed algorithm. One limitation of JESP-NE is that the quality of the obtained joint policy depends on the predefined default policy, which is used for a branch with zero probability. Let us show an example. Assume there are two agents i and j, the finite horizon T = 2, the listen cost is c = 30, and the tiger cost is d = 40. The initial policy is: first Listen, then OpenLeft when HR is observed, OpenRight when HL is observed, and OpenLeft when Reset is observed. First, we calculate the best response for agent j as: first Listen, then OpenLeft when HR is observed, and OpenRight when HL is observed. In this case, observation Reset has zero probability. In such a case, JESP-NE simply assigns some default policy in this branch. Assume the default policy is always Listen. Next we calculate the best response for agent i as: first Listen, then OpenLeft when HR is observed,
OpenRight when HL is observed, and Listen when Reset is observed. This joint policy is a Nash equilibrium. On the other hand, in JESP-TPE, when calculating the best response of agent j, we assume agent i might take an action that is not specified in the policy with a small probability. Thus, observation Reset has non-zero probability. Then, the best response becomes: first Listen, then OpenLeft when HR is observed, OpenRight when HL is observed, and OpenLeft when Reset is observed. Next we calculate the best response for agent i as: first OpenLeft (or OpenRight ), then OpenLeft when HR is observed, OpenRight when HL is observed, and OpenLeft when Reset is observed. Here, since agent j chooses OpenLeft if Reset is observed, the expected utility improves if agent i performs OpenLeft (or OpenRight ) at t = 1. Then the best response of agent j becomes: first OpenLeft, then OpenLeft when HR is observed, OpenRight when HL is observed, and OpenLeft when Reset is observed. This is a trembling-hand perfect equilibrium, as well as a Nash equilibrium. As shown in this example, by calculating the best response even though some branch that has zero probability, JESP-TPE can converge to a better joint policy than JESP-NE. Algorithm 4 computes the expected reward for an action, and Algorithm 5 updates the state. Both functions consider the possibility that the other agents may make a mistake and set probabilities which the other agents deviate from the policy to /(|Aj |−1) (see lines 5 and 13). Fig. 1 shows different belief states (and possible transitions among them) for agent i in the tiger domain. In this notation, (SL, (HR)), which indicates the current world state, is SL, while agent j has observed (HR). JESP-TPE explores the combinations of actions, observations, and histories that the JESP-NE does not consider under the assumption that no agent makes a mistake. Therefore, JESP-TPE can examine policies with belief states that JESP-NE no longer examines (see Fig. 1). For example, JESP-TPE calculates the probability of 0.5 of episodes 1 (SL(Reset)) and (SR(Reset)) on belief state B i , while JESP-NE calculates 0, as shown in Fig. 1, because agent j will probably choose OpenLeft or OpenRight, although the policy leads j to choose Listen. As a result, JESP-TPE
Multiagent Planning with Trembling-Hand Perfect Equilibrium
21
Algorithm 4. GETVALUEACTION-TPE (B t , ai , b, πj , t, T ) δj ← πj (b) value ← 0 for all aj ∈ Aj , et = st , ω j s.t. B t (et ) > 0 do if aj == δj (ω j , T − t) then + value ← B t (st , ω j ) · R(st , ai , aj ) · (1 − ) else + value ← B t (st , ω j ) · R(st , ai , aj ) · |Aj1|−1 for all ωi ∈ Ωi do B t+1 ← U P DAT E(B t , a, ωi , δj ) prob ← 0 for all aj ∈ Aj , st , et+1 = st+1 , ω j s.t. B t+1 (et+1 ) > 0 do if aj == δj (ω j , T − t) then + ← B t (st , ω j ) · P (st , a1 , a2 , st+1 ) · O1 (st+1 , a1 , a2 , ω1 ) · prob t+1 O2 (s , a1 , a2 , ω2 ) · (1 − ) 14: else + ← B t (st , ω j ) · P (st , a1 , a2 , st+1 ) · O1 (st+1 , a1 , a2 , ω1 ) · 15: prob t+1 O2 (s , a1 , a2 , ω2 ) · |Aj1|−1
1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13:
+
16: value ← prob · GET V ALU E(B t+1 , t + 1) 17: return value
Algorithm 5. UPDATE-TPE (B t , ai , ωi , πj ) 1: for all et+1 do i 2: B t+1 (et+1 ) ← 0 3: for all aj ∈ Aj , st ∈ S do 4: if aj == δj (ω j ) then + 5: B t+1 (et+1 ) ← B t (et ) · P (st , ai , . . . , aj , st+1 ) · O(st , ai , . . . , aj , ω) · 1 − 6: else + 7: B t+1 (et+1 ) ← B t (et ) · P (st , ai , . . . , aj , st+1 ) · O(st , ai , . . . , aj , ω) · |Aj1|−1 8: normalize B t+1 9: return B t+1
searches for the policies on all belief states, while JESP-NE only searches for the 1 policies on Bi1 and B i . Notice that Algorithm 4 defines value function Vtai that represents the expected reward that the agents will receive when agent i follows action ai at the t-th step. Also, Algorithm 5 defines agent i’s belief state Bit+1 at the t+1-st step when i chooses action ai and receives observation ωit+1 at the t-th step.
6
Experimental Results
In this section, we perform an empirical comparison of the JESP-NE and -TPE described in Sections 4 and 5 using the tiger scenario in terms of the expected reward of the obtained policies (see Section 3). We ran the algorithm for an
22
Y. Yabu, M. Yokoo, and A. Iwasaki 20
expected value
15 10 5 0 -5 -10 -15
setting 1
setting 2
setting 3
Fig. 2. Expected Reward obtained by the JESP-NE and the JESP-TPE
Fig. 3. Expected reward for listen cost c with tiger cost d = 14
initial state chosen randomly, and the initial/default policy is selected as Listen for all states. Probability , with which an agent makes a mistake, is set to a small value, i.e., 1.0E − 13 so that the expected reward is not affected. Fig. 2 shows the expected reward of the two JESPs for three different settings where the tiger cost is 20. When we ran the algorithms with a listen cost of 8 and time horizon is 4 (setting 1), both expected rewards are 2.25. Next, we increase the listen cost to 14 (setting 2). JESP-TPE outperforms JESP-NE, i.e., the expected rewards are 10.0 and −9.70. Then we increase the finite horizon to 5, keeping the listen cost of 14 (setting 3). JESP-TPE again outperforms JESP-NE, i.e., the expected rewards are 15.0 and −4.75. Let us discuss why our proposed algorithm outperforms the JESP-NE in terms of the expected reward under settings 2 and 3. When both algorithms compute the policy at the first iteration, action Listen at the first step is the best action. In JESP-NE, state B 1i is never reached on Fig. 1 (JESP-TPE changes the policy on Bi0 from Listen to either OpenLeft or OpenRight. However, in setting 1, the expected reward in JESP-TPE is identical to JESPNE. Recall that Listen is the best action at the first step. Thus, since the policies 1 on B i are already optimized, the policy does not change even in JESP-TPE. As a result, the expected reward in both algorithms is identical. Fig. 3 shows the expected reward for a variety of listen costs, i.e., for 1 < c < d at T = 4 and d = 14. We achieve qualitatively similar results when the tiger cost is changed in the range of [1, 100]. Thus, we fix it to d = 14. For c ≤ 6 and c ≥ 11, the expected reward in both algorithms is identical, because they converge at the same equilibrium. For 7 ≤ c ≤ 10, JESP-TPE significantly outperforms the JESP-NE, since they converge at different equilibria. In summary, our results suggest that, when the listen cost c is about half of tiger cost d, JESP-TPE is likely to outperform JESP-NE. On the other hand, when c is very small or very large, the expected rewards of both algorithms are about the same. Furthermore, we examined the trajectories of JESP-NE and -TPE until they converge to an equilibrium. Figs. 4-6 show the expected reward for each iteration. The trajectories in Fig. 4 are exactly identical, since both algorithms converge to the same equilibrium. Fig. 5 shows that JESP-TPE outperforms JESP-NE
Multiagent Planning with Trembling-Hand Perfect Equilibrium 20
23
20 10
10 expected value
expected value
0 0 -10 -20
-10 -20 -30 -40 -50
jesp-ne jesp-tpe
-30
jesp-ne jesp-tpe
-60 0
1
2 3 number of iteration
4
5
Fig. 4. Expected reward for number of iterations with T = 4, c = 8, and d = 20
0
1
2
3 4 5 number of iteration
6
7
Fig. 5. Expected reward for number of iterations with T = 4, c = 14, and d = 20
20 10
expected value
0 -10 -20 -30 -40 -50 -60
jesp-ne jesp-tpe
-70 0
1
2
3 4 5 6 number of iteration
7
8
9
Fig. 6. Expected reward for number of iterations with T = 5, c = 14, and d = 20
except for the expected reward at the first iteration. JESP-NE converges to an equilibrium at the second iteration, while JESP-TPE does so at the fourth iteration and reaches a better joint policy. Fig. 6 shows a qualitatively similar result to Fig. 5 in setting 3, i.e., the expected reward starts at the same value, and as the iteration continues, JESP-TPE eventually outperforms the JESP-NE.
7
Conclusion
Multiagent POMDPs provide a rich framework to model uncertainties and utilities in complex multiagent domains. Since the problem of finding an optimal joint policy for a multiagent POMDP is NEXP-Complete, we need to sacrifice optimality to some extent. In this paper, we proposed a method that finds a locally optimal joint policy that corresponds to a Trembling-hand Perfect Equilibrium (JESP-TPE), while the existing approach, JESP-NE, finds a joint policy that corresponds to a Nash equilibrium. TPE is stable even if we assume that an agent might make a mistake in selecting its action with small probability, and it is less dependent on the quality of a predefined default policy. The experimental
24
Y. Yabu, M. Yokoo, and A. Iwasaki
results showed that JESP-TPE outperforms JESP-NE in settings where the default policy is not good enough. In future works, we would like to examine our algorithm in more practical real-world problems.
References 1. Beard, R.W., McLain, T.W.: Multiple uav cooperative search under collision avoidance and limited range communication constraints. In: Proceedings of the 42nd Conference Decision and Control, pp. 25–30. IEEE, Los Alamitos (2003) 2. Nair, R., Tambe, M.: Hybrid BDI-POMDP framework for multiagent teaming. Journal of Artificial Intelligence Research 17, 171–228 (2002) 3. Lesser, V., Ortiz, C., Tambe, M.: Distributed Sensor Networks: A Multiagent Perspective. Kluwer, Dordrecht (2003) 4. Xuan, P., Lesser, V., Zilberstein, S.: Communication decisions in Multiagent cooperation. In: Proceedings of the Fifth International Conference on Autonomous Agents, pp. 616–623 (2001) 5. Goldman, C.V., Zilberstein, S.: Optimizing information exchange in cooperative multi-agent systems. In: Proceedings of the Second International Joint Conference on Agents and Multiagent Systems (AAMAS 2003), pp. 137–144 (2003) 6. Nair, R., Tambe, M., Marsella, S.: Role allocation and reallocation in multiagent teams: Towards a practical analysis. In: Proceedings of the Second International Joint Conference on Agents and Multiagent Systems (AAMAS 2003), pp. 552–559 (2003) 7. Bernstein, D.S., Zilberstein, S., Immerman, N.: The complexity of decentralized control of markov decision processes. In: Proceedings of the 16th Conference on Uncertainty in Artificial Intelligence (UAI 2000), pp. 32–37 (2000) 8. Nair, R., Roth, M., Yokoo, M., Tambe, M.: Communication for improving policy computation in distributed pomdps. In: Proceedings of the Third International Joint Conference on Agents and Multiagent Systems (AAMAS 2004), pp. 1098– 1105 (2004) 9. Selten, R.: Reexamination of the perfectness concept for equilibrium points in extensive games. International Journal of Game Theory 4, 25–55 (1975) 10. Pynadath, D.V., Tambe, M.: The communicative multiagent team decision problem: Analyzing teamwork theories and models. Journal of Artificial Intelligence Research 16, 389–423 (2002) 11. Mas-Colell, A., Whinston, M.D., Green, J.R.: Microeconomic Theory. Oxford University Press, Oxford (1995)
MAGEFRAME: A Modular Agent Framework to Support Various Communication Schemas Based on a Self-embedding Algorithm Quintin J. Balsdon and Elizabeth M. Ehlers The University of Joahnnesburg: Academy for Information Technology Tel.: +27 11 559 2847; Fax: +27 11 559 2138
Abstract. The communication techniques used amongst agents in a system is critical to the effectiveness and success of collaboration based actions. Agents which fail to store or communicate information in an effective uniform manner will not be able to accomplish their goals. The aim of the paper is to determine the particular difficulties relating to agent information representation and reconstruction. In addition the paper will discuss whether or not it is profitable for agents to be equipped with the ability to modify their communication style. Keywords: Agents, Communication, Collaboration, Agent Self-Modification, Embeddable Components, Agent Framework.
2 Communication Concerns The problems that may be encountered during the communication process and the methods of communication will be discussed. Further communication concerns discussed are the composition of messages, the integration of information and the role of the environment. 2.1 Factors Which Play a Significant Role in the Interpretation of a Message A message is a single formal communication that requires the interpretation of the entity that is involved in its reception. The process of communication is identified by Russell and Norvig as one which involves: intention; generation; synthesis; perception (of the message as a communication unit); analysis; disambiguation and incorporation [1]. Each of the identified stages, should they be hindered, may lead to the modification or corruption of the message that is intended to be sent. The process of successful message transfer is understood to be a period in time after a message has been initiated by a sender and received by the intended recipient, which regards the information transferred in exactly the same manner as the information is regarded by the original sender1. Both sender and receiver have significant roles in the successful transmission of a message. The sender involves itself with intention, generation and synthesis, while the receiver is engaged in perception analysis, disambiguation and incorporation [1]. The sender is responsible for three stages in which erroneous communication may occur. In the intentional stage, the timing or purpose of the message may effect the message adversely. In order to avoid this problem, the receiving agent could perform a validity check on the message, which will be aided in sending the actual intentions of the message as well. The generation of the message effects the messages understandability on behalf of the receiving agent. A poorly generated message cannot be clearly disambiguated on the other end. Message synthesis involves the sender and the environment. A message must pass through some medium inside the environment to reach the destined receiver agent. The receiving agent, upon the delivery of a message, must first separate the message from the medium in which it is transported. After this, the message must have all ambiguity removed, preferably using the context the message was sent in. Should the receptor agent successfully receive a message that has more than one possible meaning, the best context that can be applied must be used. Rather than have agents calculate what the context of a message should be, some contextual evidence may be provided on behalf of the sender. However, there are some problems with this procedure: the amount of contextual evidence which should be provided, considering the amount of perception the sender has at the specific time of message generation, without overloading the recipient; the determination of the correct evidence to provide the recipient, considering the context as applied to the recipient. These problems were 1
This is not to say that the receiver will act in the same way regarding the received information, but rather that the information will mean the same to both agents contextually. Each agent may perform different actions, but their premise will be the same information with the same implications.
MAGEFRAME: A Modular Agent Framework
27
identified by J. Kant and S. Thiriot, in their paper entitled “Modeling one human decision maker with a multi-agent system: the CODAGE approach”, when they stated “due to its bounded rationality, the decision maker cannot represent the whole world” [2]. Information received may have many differing effects on the agent. The receptor agent must determine whether the information is to be discarded, or if the information must be included somehow. The beliefs must be consistent with all information that is accepted by the agent, making the communication of a belief vitally important. Any other kind of information sent must be consistent with the beliefs, or if the communicated belief is found to be true, all relating beliefs must be negated. The agent will have to determine whether or not it should negate all relating BDI (belief, desire and intention) elements. Message incorporation is an important aspect of the agent communication cycle and will be discussed in further detail. Communication in complex environments with agents differing in complexity or design is prone to errors along the entire pipeline of communication. Unless these agents are programmed with the ability to modify their behaviour in a manner which avoids these problems, the systems that these agents operate in will continue to require both simple designs and environments. These issues must be addressed individually, but in the most concise manner possible in order to maintain an efficient environment. 2.2 Communication Methodology Many differing communications styles have been proposed for agents that are required to interact in a system. Since the medium of the environment in question is software-based, the medium that the agents will utilize to communicate in will also be established in the software medium. In an environment where the activities of agents may be distributed and various communications maybe used, agents may fall into one of two categories: the receiverinitiated and the sender-initiated [3]. In the receiver-initiated method, the agent will confer with other respective agents as to whether they have a message to send. In this type of communication strategy the receiver is introduced as the actual initiator in the communication process, although the sending (‘intentioning’) agent remains the originator. B. Wilkinson and M. Allen identify this method as working well under high system load [3]. The alternative involves the originating agent selecting the agents which ought to receive the message and proceeds in ensuring that those agents receive the message [3]. Agents may employ a strategy where there is no intermediary mechanism. This would typically allow an agent to bypass the environment, due to the fact that the information would travel from agent to agent, regardless of the medium(s) that inhabit the environment. While it may be useful to agents with the same design or easily compatible, the problem would remain that this type of communication is one to one. If an agent would communicate the same message to a group of agents, then the agent would, in the best case scenario, prepare a message and deliver the same message to each member of that group one by one. In the worst case scenario, it may have to prepare a distinctive message for each member and deliver it on a one to one basis. While this technique would work in various systems, the agent may spend more time communicating in complex multi-agent environments than attending to its goals.
Q.J. Balsdon and E.M. Ehlers
Communication Flow
28
Sender
Receiver
Message
Fig. 1. The difference between a blackboard structure and a chain structure
One proposed communication method is the blackboard [4]. A blackboard is typically a sender initiated communication strategy. The main problem with a blackboard is synchronization. An agent may post a message on the blackboard, only for the other agents to respond to that message out of time. A solution to this problem would be to have the agents subscribe to the blackboard, so that when a message is posted the blackboard sends a small and identical message to all the subscribing agents. This centralised form of communication is best utilized in environments where there are many agents that are grouped together, to ensure that message consistency is maintained. Another means of message dissemination would be to use chains. A message chain could be formed in the communicating agent groups so that the information is sent out from the sender to the other agents in the group. This would eliminate the use of a blackboard structure and aid the originator agent in such a manner that it would not send the same copy to every receiving agent. A message chain is most optimal in agent groups comprised of a few agents where messages are passed quickly. The problem with the structure is that if one agent fails to send the message then all the agents along its chain will fail, although this method could be enhanced to utilize graph structures rather than tree structures. There are several methods for sending messages over a medium in a software environment. The two examples given, a blackboard and a chain, as in figure 1 above, illustrate the manners in which a medium may be provided. The latter variety is to create a specific medium inside the environment that is dedicated to message delivery, while the former uses the agents in some mode as the means themselves. 2.3 Message Composition While delivery has been the focus of most of the discussion up to this point, the composition of a message must be considered. The content of the message itself must be distinguished from the composition of the message. The composition of the message shall refer to that which makes the message an entirety, in other words, everything that is transmitted from one agent to another. The content of a message is used to refer to that which is the focal point, or the purpose behind the intention of the message
MAGEFRAME: A Modular Agent Framework
29
generation. The content of the message therefore forms a part of the composition of the message. Since different agents of varying complexity or design must be able to communicate, there must be some base level descriptor of the manner in which the elements that comprise the composition are identified. The structure of the messages must be formalized, so that if agents do not use certain components, the agents will be able to communicate with each other. A tuple2 will therefore be utilized to represent this structure. Each communicative method must therefore be able to subscribe to the utilization of a tuple. This does not inhibit the freedom of communication that each agent may exhibit in the environment. While all agents will be employing tuples as wrappers, the contents of each compositional element will remain in the originator agent’s communicative language. This does not inhibit the ‘language’ of each agent, but ensures that each ‘language’ is ‘understandable.’ The ontology contained within the tuple must be composed of elements that describe the objects in the environment and elements that provide details on the manner in which these objects relate to each other. In this manner any agent may break down these elements to interpret the message. As stated above, the intention of a message must be communicated in order that the receiver agent may perform a check that would indicate the message’s validity. Suppose that an agent, A, wishes a situation to change and sends a message to an agent, B, to stop its current actions and aid agent A. Agent B may receive the message after an external force (a force other than agent B) has performed an act in such a manner that the act agent A is requesting of B is no longer necessary. Agent B may examine the intention behind the message and perceive that the intention is no longer valid. Therefore, agent B is not required to consider the message and may continue without regarding the message, which would be the most efficacious use of computer cycles. As previously discussed, an agent may be required to insert multiple intentions into a single message. It must be noted that the intentions of the message must not contradict the intentions of the originator agent. However, the message intentions may not exactly reflect the agents’ intentions as described by the BDI model. The intention of the message may be left blank, in order that an intention is being communicated. There are several purposes for sending a message, as identified by [1, 6]. Table 1 below describes the differing incentives for messages and the effected BDI element. The message incentive does not denote a complete set of intentions for the recipient. The intention of the message must be communicated in terms of the current state of the environment. While it would be effective for an agent to send the incentive, perhaps to aid in disambiguation or consistency analysis, it would not be as important as the contextual intentions based on the environment. A further component in the message composition is the element of communication that has an effect on the beliefs, desires or intentions of the agent recipient. The content, as opposed to the intentions, is a single entity in the communication. It is not advisable that a message contains various sets of content, since some agents may share some beliefs, desires or intentions and differ on others. The agent must be able to determine whether a message is to be incorporated or discarded based on a combination of the message’s intention and the content of the message. 2
“A tuple is formed by taking a list of two or more expressions of any types, separating them by commas, and surrounding them by round parenthesis” [5].
30
Q.J. Balsdon and E.M. Ehlers
Message Incentive Assertive Directive Commitment
Permission
Declaration
Expression
Description Information which pertains to the environment Appeal to the receiver to respond in the form of action An type of acknowledgement which is the response to a directive. Indicates that an agent is dedicated to that action. An agent which has authority over a group may have to indicate that an action may or may not be performed An agent may communicate a formulation that effects the manner in which other agents operate A message to communicate feelings
BDI element that is effected Beliefs Desires Intentions
Intentions
Beliefs
None
Table 1. The different categories of message intentions and their effect on the receiver
A message tuple would look as follows: (i0 ...im ), ( j0 ... jn ), (c0 ...c p ) Where i is indicative of the incentive constants for the message, j is the list of contextual intentions and c depicts the contents of the message. Each element in the tuple would be represented in the communicating agent’s chosen communication schema. 2.4 Information Integration: The Manner in Which New Information Is Assimilated If the information received contradicts the beliefs of the agent, the agent should consider whether it is going to continue with its current belief system, or negate the belief and accept the information. An agent whose utility is found in computational efficiency would secure a higher utility by rejecting (negating) the information. This is due to the fact that it requires a greater amount of computation effort to not only modify its beliefs, but to modify all existing information stored in the database consistent with the belief. This may be an action resulting in long term utility. If the agent has direct visible evidence, which may be communicated in the form of intentions or by what is observed in the environment, that its beliefs are incorrect, it would be better suited to the agent to modify itself than attempt to modify the world based on beliefs that are contrary to the world. Agents ought to not have false beliefs due to the fact that beliefs are based on perceptions of the world. However, due to the contextual variance that may be obtained by submitting agents to very dynamic worlds, the fact that agents may be created with inconsistent beliefs remains a possibility.
MAGEFRAME: A Modular Agent Framework
31
6
1 14
7
5
2 5
13
15
10
2
9
4
4 1 3
6
8
16 11
3
7
Attributes which are the building blocks of constraints
8
12
Constraints which connect attributes by means of relations
18 17
Propositions connect constraints by adequate sets of connectives
A system of contradiction – free propositions
Fig. 2. A system of two propositions being introduced to a third
An agent accepting new information must not accept (or assimilate) information that contradicts with its own BDI representations. If the beliefs are to be modified, the agent would have to ensure that the correct course of action is being taken, as discussed above. If the desires or intentions require modification, the new beliefs and desires must not contradict with the agents existing desires and intentions. This means BDI elements of agents must be either exclusive or consistent. If two elements are exclusive, it means that the two elements have nothing to do with each other; the agent may achieve both elements without affecting the other, regardless of the order of execution. If two elements are to be consistent, the elements have common components, but the achievement of one will aid the achievement of the other. Consistency may be examined by combining the agent’s existing constraints (see figure 2) and the new proposition into a conjunctive formula and checking the formula for contradictions. However, due to the fact that adequate sets of connectives [7] allow for a formula to be represented in several forms, this method, although simple, may become unnecessarily time consuming. By performing consistency checking using a conjunction, each new proposition will have to be checked for consistency against all existing propositions. This could become computationally intensive when there are many existing propositions. It is proposed that formalism be defined to reduce the amount of checks that will be performed on the set of propositions. Figure 2 depicts a possible situation in the information incorporation process.
32
Q.J. Balsdon and E.M. Ehlers
Logic L has the Robinson property (or, equivalently in L the Robinson consistency
τ ,τ 1 and τ 2 with τ = τ 1 ∩ τ 2 , if T1 and T2 are consistent extensions of T respectively in type τ 1 and τ 2 , and T is complete in τ , then T1 ∩ T2 is consistent in τ 1 ∪ τ 2 . theorem holds) iff for any theories T, T1 and T2 and types
Definition 1: The Robinson property
Consider a contradiction free system of propositions P1…Pi which are constructed by adequate sets of connectives and contain a subset of constraints c1…cj. Suppose there is a proposition Q, containing the constraints q1…qk, which is constructed by the same sets of connectives as P1…Pi. In order to include Q where Q is not already within the system, in such a manner that there are no contradictions in the system, it is only necessary to demonstrate that Q does not introduce a contradiction for those constraints such that ql and cm ...cn are consistent where ql contains the same attributes, a1…ap, as ( m< n )
cm ...cn
. It follows that constraints which consist of attributes that are not contained within the system will have no effect once they are added. Robinsons consistency theorem, as in Definition 1, supports this model [7]. The message may be rejected by the recipient agent based on one instance of disagreement with any of the schema elements. If the agent does not share: a belief consistent with the incentives; a desire that is consistent with the contextual intentions or the contents are inconsistent with the current intentions of the agent, the recipient agent may either reject the information and continue, or assimilate the negation of the content. ( m< n )
2.5 The Role of the Environment The environment is a crucial aspect of the agents which have dominion over it. Agents must be able to fully function inside their environment in order for them to be efficient. Agents of differing complexities may be equipped in various ways regarding their communication components. They may have a standard component which must be used, or they may be able to strip itself of some components and equip different components. The environment must be able to support the variable agent frameworks that may be found inside itself. The environment itself is not responsible for managing communication, but rather making sure that communication is possible. Therefore agents must be able to use the components of the other agents to communicate amongst one another. The environment must provide a repository that agents may utilize to encode or decode messages sent by other agents. Agents in the system are then required to register their components at the environment communication repository. While using a communication repository, an agent will have to determine whether it is best to communicate in the receptors communicative language, or to require the receiver of the message to translate the message. If both the agents involved can equip any communication component, then the sender agent may conclude in either manner. However, a situation may possibly arise where either or both agents are unable to strip
MAGEFRAME: A Modular Agent Framework
33
and equip components. A component repository could be in the form of a complex agent that may perform the translation for the agents, working on a mapping table. A mapping table would simply have the format of the sent message and convert it to the format of the receiver.
3 The Modular AGEnt FRAMEwork (MAGEFRAME) The agent framework, called MAGEFRAME, enables an agent to embed different communication modules within its structure. The agent model as described by figure 3 is two-fold: each element in the agent process (indicated in a) will be equipped with the same hierarchical structure (indicated by b). The sensors and actuators are both involved in the agent’s proceedings as they are directly related to the environment. With these components the agent is able to formulate and pursue its beliefs, intentions and desires. The performance element is the core component of the agent structure. The agent will have its overall beliefs, intentions and desires stored in this component. The performance element is responsible for ensuring that the minimum utility of the agent is achieved. The problem generator exists to perceive possible undesirable action patterns and suggest alternative actions [1]. The agent may ‘learn’, in a primitive manner, by pattern recognition that certain actions always lead to undesirable or least satisfactory results. In this manner the agent may be able to discover a method of achieving its maximum utility. The communication element of the agent is designed to deal with the external communications being sent and received through the media within the environment. The component utilizes a structure that allows the embedding of communication components. The embedded communication component within the agent’s communication element performs the communication with external agents, while the communication components in all the other agent components manage internal communication. This component will either make direct changes to the connections inside the environment or make connections directly with other agents in the environment; this decision is dependant on the environment. The reactive components ensure that the agent does not allow an unsafe situation to persist. The reactive components will be provided by the environment in a table format that describes the manner in which an agent is to act depending on pre-defined conditions. The agent will not spend computational cycles interpreting inputs that are listed on the table from the environment, for the sake of safety. The inter-agent communication element describes the manner in which agents are to communicate internally. The agent represents information in a standard fashion and sends it directly to the appropriate component. While the performance element concerns itself with utility, the communication element will determine the worthiness of information that is sent inside the agent. The BDI model contained in each agent module relates to the elements which are most important to that module. In the performance element, there are very few to no beliefs, as it is the beliefs that will inhibit the agent of considering abandoning certain beliefs. The performance element contains all the beliefs that the agent carries. The
34
Q.J. Balsdon and E.M. Ehlers (a)
Performance Standard
Sensors
Performance Element
Problem generator
tion ges
Environment
Sug
ns isio De c
s
Agent
Communication Element
Actuators
(b) Beliefs, Desires, and Intentions Executive layer
Higher
Inter-Agent Communication Element Reactive layer
Built-In Reactive Components
Lower
Fig. 3. The MAGEFRAME agent model
communication element carries some of the beliefs of the agent, with which it would determine what may be communicated. In relation to the BDI model, the central component is the performance generator. The performance generator must determine if a BDI element should be discarded or modified. The communication element will then perform the modification. The MAGEFRAME agent framework differs from its predecessors [1,8] due to the combination of two structures. The agent framework (indicated by figure 3a) supports the lifecycle of the agent. Internally a hierarchical subsumption model (indicated by figure 3b) provides each module in the framework with a standardized process of behaviour that must be followed. The manner of behaviour may be modified by the different embeddable components within the internal structure. However, embedding components is not necessary; the agent is only required to register its communication modules within the environment.
4 Communication Schemas While an ontology that is flexible enough to be complete in all situations is desirable, this cannot be the final solution. This is due to the fact that all agents may not be equipped with such an ontology, while other agents may not be able to maintain a complete ontology. In their paper entitled “ANEMONE: an effective minimal ontology negotiation environment”, J. van Diggelen, et al, state the reason for minimizing an ontology “An ontology that is minimal in size is not bulky and slowly process-able due to superfluous concepts” [9].
MAGEFRAME: A Modular Agent Framework
35
Agents utilizing MAGEFRAME are equipped to communicate with other agents that are of the same design and complexity. However, if an agent with a different design or complexity is introduced to the system, all the agents in the system must be able to determine the meaning and context of the messages in the system.
5 A Test Environment The proposed test environment is a ‘lecturer-student world’ where students must consult various sources, for example lecturers and textbooks, to achieve the goal of ‘passing.’ The alternative to ‘passing’ is ‘failing.’ This environment is a more complicated environment based on S. Thrun’s ‘Heaven-hell’ world [10]. The environment will use a similar ‘success policy’ as to the ‘Heaven-hell’ world, meaning that the agent will always reach its goal, no matter how inefficient the path found may be [10]. The agent may ‘fail’ several times, however it will be permitted to continue through the environment until it achieves a ‘pass’ status. The maximum utility for the agent would be to reach the state of ‘distinction.’ The agents in the environment, namely the textbook, lecturer and lecture notes, as well as other types of agents communicate in differing styles to one another. Their communication components must be registered in the environment and another ‘converter’ (tutor) agent will exist as well. The tutor agent translates from one language to another for agents that cannot perform the conversion themselves, although it is not exclusively for this type of agent. More complex agents may determine the usage of the tutor agent is more efficacious than performing the translation themselves.
6 Conclusion A solution to the communication problem between agents has been proposed. Different issues in the communication process have been discussed, as well as a methodology of communication. Agents must integrate information with great care, as the consistency of their BDI elements is very important. The environment may also play a role in the success or failure of a communication. The agent framework MAGEFRAME has been proposed that not only allows agents of differing complexity to communicate, but may simply be an intermediary communicator for any two agents. Agents may not always use complete language schemas because they tend to be computationally expensive. A test environment for multi-complexity-agent communication and interaction has been identified.
References 1. Russell, S., Norvig, P.: Artificial Intelligence: A Modern Approach. Alan R. Apt, Pearson Education, Inc., London (2003) 2. Kant, J., Thiriot, S.: Modeling one human decision maker with a multi-agent system: the CODAGE approach. In: AAMAS 2006: Proceedings of the Fifth International Joint Conference on Autonomous Agents and Multiagent Systems, Hakodate, Japan, pp. 50–57. ACM Press, New York (2006)
36
Q.J. Balsdon and E.M. Ehlers
3. Wilkinson, B., Allen, M.: Parallel Programming: Techniques and Applications Using Networked Workstations and Parallel Computers. K. Hargett. Pearson Prentice Hall, Upper Saddle River (2005) 4. Gamatié, A., Gautier, T., Le Guernic, P., Talpin, J.P.: Polychronous design of embedded real-time applications. In: ACM Trans. Softw. Eng. Methodol., vol. 16(2), p. 9. ACM Press, New York (2007), http://0-doi.acm.org.raulib.rau.ac.za:80/ 10.1145/1217295.1217298 5. Ullman, J.D.: Elements of ML Programming. In: Horton, M. (ed.). Prentice Hall, Upper Saddle River (1998) 6. Singh, M.P.: Agent Communication Languages: Rethinking the Principles. In: Computer, vol. 31(12), pp. 40–47. IEEE Computer Society Press, Los Alamitos (1998), http://dx.doi.org/10.1109/2.735849 7. Mundici, D.: Robinson’s Consistency Theorem in Soft Model Theory. In: Transactions of the American Mathematical Society, vol. 263(1), pp. 231–241 (1981) (2007/3/12), http://links.jstor.org/sici?sici=00029947%28198101%29263%3A1%3C231%3ARCTISM%3E2.0.CO%3B2-O 8. Huhns, M.N., Rose, J.R., Roy, S.S., Turkett Jr., W.H.: An Agent Architecture for LongTerm Robustness. In: Proceedings of the First International Joint Conference on Autonomous Agents and Multiagent Systems: part 3, vol. 3, pp. 1149–1156. ACM Press, New York (2002) 9. van Diggelen, J., Beun, R., Dignum, F., van Eijk, R.M., Meyer, J.: ANEMONE: an effective minimal ontology negotiation environment. In: AAMAS 2006: Proceedings of the Fifth International Joint Conference on Autonomous Agents and Multiagent Systems, Hakodate, Japan, pp. 899–906. ACM Press, New York (2006) 10. Thrun, S.: Monte Carlo POMDPs. In: Solla, S.A., Leen, T.K., Müller, K.R. (eds.) Advances in Neural Information Processing Systems 12, pp. 1064–1070. MIT Press, Cambridge (2000)
Using Multiagent System to Build Structural Earth Model Beiting Zhu1, Zahia Guessoum2, Michel Perrin3, Bertrand Braunschweig1, Pierre Fery-Forgues1, and Jean-François Rainaud1 1
Institut Français du Pétrole, 1 & 4, av. de Bois-Préau, 92852 Rueil-Malmaison, France {Beiting.Zhu,Bertrand.Braunschweig,J-Francois.Rainaud, Pierre.Fery-Forgues}@ifp.fr 2 Laboratoire d’Informatique de Paris 6, 104 av. du président Kennedy, 75016 Paris, France [email protected] 3 Ecole des Mines de Paris, 60 boulevard Saint-Michel, 75272 Paris, France [email protected]
Abstract. Mainly based on seismic data, structural earth models are commonly used for oil & gas reservoir engineering studies. The geometry and the topology of a structural earth model strictly depend on the geological characteristics of its various surfaces. These characteristics constitute the geologists' interpretations of the earth model to be built. In the 3D modeling applications currently used in industry, these interpretations are often registered after the completion of the structural earth model. Therefore further changes in the interpretations can hardly be introduced without reconstructing the structural model from scratch. This situation is dramatic since the only true geometrical information concerning subsurface is that issued from drillings, which daily bring new data. For the above reasons, there is a great interest in providing computer-aided tool for validating and improving geological interpretations to facilitate the geologists’ modeling work. This paper presents a distributed problem solving approach, which enables geologists to quickly validate and improve structural earth models by considering new drilling data. This approach is based on a selforganizing multiagent system and on domain-specific expertise. Keywords: Cognitive agent model, multiagent system, distributed problemsolving, geological interpretation, structural earth modeling.
Earth model initialization and revision are the results of several experts’ cooperation. These experts come from various disciplines, such as: geophysics, structural geology, drilling engineering, etc. They update the earth models by taking into account the drilling data brought daily. Such data, which usually gives rise to modifications of experts' current interpretations, is unpredictable and is imported at any time while the structural model is evolving. Therefore, revising the current model by introducing the new data and the consequent changes becomes an important but hard work. However, in the modeling applications currently used in oil & gas industry, there is no means of registering and supporting the various successive interpretations. Therefore, it is difficult to update the current earth model, which must often be rebuilt from scratch. This task is as more difficult as the earth models' interpretations become numerous. So, it is desirable to study computer-aided solutions, which would provide a help to the modelers for improving the earth model in a semi-automatic way, so as to easily take into account the daily additional information provided by the advancing drillings. This paper introduces a distributed solution of the above issue, which is realized by a Multi-Agents System based on the eco-problem solving methodology [4]. In ecoproblem solving approach, the problem is decomposed into a set of local subproblems. The problem-solving issue is delegated to a population of independent reactive agents. Their interaction progressively induces a coordination of their behaviors, which finally leads to a stable collective self-organization. The solution of the problem is the direct consequence of this stable organization. Such reactive approach has been applied to resolve complex problems as [2], [3], [8]. Its simplicity allows solving complex problems without predicting an overall strategy. In this principle, we propose to associate an agent to one geological surface in a structural earth model. Each agent tries to maintain its surface's coherent geological interpretations. For efficiently achieving its tasks, it is built to be cognitive. This paper is organized as follows. Section 2 introduces the necessary geological background, and details the problem considered in this paper. Section 3 presents our distributed problem-solving approach based on a multiagent model. Section 4 presents a cognitive agent model, which is composed of a behavior model, a communication model, and a knowledge model. Section 5 describes the experimentation operated on two simulated examples defined by domain experts. Finally, some conclusions are presented in Section 6.
2 Structural Earth Modeling Background In this section we firstly examine how structural earth models are presently built and how their building can be partly automated. We then define at which conditions such models can be improved in view of additional drilling information. 2.1 Structural Earth Model Construction Earth models related to oil & gas exploration and exploitation basically rest on seismic data. Additional information is provided by the drilling survey data collected along well bore trajectories during reservoir exploration. In order to approximate the modeling result to real nature, earth models should be revised to integrate all the available drilling survey data issued from progressing drillings.
Using Multiagent System to Build Structural Earth Model
Φ1
Φ2
39
S2
S2 S2
S1 S1
G G
Fig. 1. The left side shows a structural earth model in a 3D block view. The right side shows the same model in a 2D cross-section view. In this model, we have the intersection such as the oldest "on-lap" horizon G interrupts a younger "parallel" horizon S1.
Fig. 2. A naïve example showing how geological interpretations affect the topology of two intersecting surfaces [10]
Here we talk about structural earth models, which correspond to mere assemblages of geological surfaces such as horizons and faults. For building purpose, modeler needs to define the all over surface arrangement assemblage, i.e. the topology pattern defining how each of these surfaces intersect the others (cf. Fig. 1). A structural model's topology pattern depends on the geological characteristics of its various surfaces (such as: erosional or on-lap horizons, constituent faults of a fault network), and on their relative ages. In the sketch b of Fig. 2, B is an “erosional surface” younger than A. B interrupts A. The points located at the right of the intersection between the two surfaces belong to surface B. In the sketch c of Fig. 2, A is an “on lap surface” older than B. B stops on A. The points located at the right of the intersection between the two surfaces belong to surface A. Joint work operated at IFP1 and ENSMP2 [1] has demonstrated an automatic manner to construct a structural earth model by starting from raw geological surfaces (faults or horizons), and from their geological interpretations formalized in a Geological Evolution Schema (GES) [1], [9]. Fig. 3 represents a GES view of the geological scene pictured in Fig. 1. GES is defined as a directed acyclical graph, whose individual nodes correspond to various geological surfaces. The geological characteristics are expressed by attributes putting on the nodes. The arcs linking the nodes correspond to chronological relations, such as S2 is younger than S1, or to topological relations inside a fault network, such as Φ1 stops on Φ2 (cf. Fig. 3).
1 2
IFP: Institut Français du Pétrole. ENSMP: École des Mines de Paris.
40
B. Zhu et al.
Fig. 3. A geological scene represented in a 3D block, and its associated GES E1
E2
O1 M O3 O2
Fig. 4. The next possible neighboring surfaces can be encountered, after marker M, along any drilling trajectory
2.2 The Issue of Validating and Updating Geological Structure Interpretations The results presented above indicate a new “knowledge-driven” approach for earth modeling work. Our issue is therefore to provide additional tools for facilitating the validation and update of the given geological interpretations, in accordance with new drilling data. This paper provides an agent-based approach, which is to help the geologists to validate the current interpretations expressed in a GES, and then to suggest and to realize possible modifications. The concerned drilling data refers to an ordered list of markers issued from the geophysicists' interpretation of a drilling log. Each of these markers corresponds to the intersection of a given geological surface (horizon or fault) by a drilling trajectory. The sequence of the markers thus expresses in which order the various geological surfaces are met by the drilling trajectory. Therefore, the marker sequence can be used to check step by step the chronological-topological relations existing between the geological surfaces corresponding two neighboring markers. These relations should strictly be consistent with the geological characteristics of the geological surfaces. Fig. 4 provides a geological scene consisting of a succession of sedimentary sequences, each of which is limited by erosional or on-lap surfaces. By supposing the scene corresponds to an interpreted structure, geologists assume that, along any drilling trajectory, the possible neighbor of marker M can only be located in the surfaces pictured in bolder lines on the figure. In the case that such a neighborhood cannot be validated, the earth model shown on Fig. 4 needs to be revised.
3 A Multiagent Approach A structural earth model describes an assemblage of geological surfaces. It is considered as consistent iff the topology of the all over assemblage is compatible with the
Using Multiagent System to Build Structural Earth Model
41
Fig. 5. The use case diagram describing the system requirements
geological age and the geological characteristics of each of the constituent surfaces in this earth model. Experts define these elements in geological structure interpretations. The consistency between the topology and the geological interpretations needs to be checked both locally (among all the interpretations related to one surface) and globally (among all the interpretations related to different surfaces). Therefore, the work defined to validate and to update the geological description of a full structural earth model is actually to validate, to update and to integrate the geological interpretations related to each constituent surface. In this sense, the entities in question are actually the surfaces. Therefore, it is natural to consider the problem in the unit of surface, rather than of the full earth model. Fig. 5 illustrates the system requirements in an UML use case diagram. The validation and update work needs to be operated in view of real-time drilling information. Since geological surfaces are related by numerous mutual relations, it is difficult to predict how a change concerning the interpretations of a limited part of the earth model will actually propagate throughout the whole model. It is therefore complex to propose an overall revision strategy that takes into account all possible changes in accordance with new drilling information, and permanently keeps the global consistency of the full earth model. We therefore propose to resolve this problem by using a self-organizing multiagent system. We decompose the global problem related to the full earth model into subproblems each related to a particular constituent surface. We then associate an autonomous agent with an individual surface, so as to resolve the various subproblems concerning this surface. By considering a multiagent system as a set of possible organized agents that interact in a common environment [4], we consider that one solution of the global problem will emerge from the agents’ self-organization. This distributed approach has two advantages: 1. The problem decomposition that is made allows to deal with smaller and more manageable sub-problems in a relative isolated manner, to limit knowledge acquisition and knowledge modeling scope, and to consequently deal with less complex sub-problems [7].
42
B. Zhu et al.
2. Each autonomous agent associated to an individual surface has the goal of validating and updating the geological interpretations regarding this surface. The agent’s autonomic and proactive features allow to automate the decision-making of choosing an appropriate behavior, so as to react to unpredictable situations raised by real-time drilling information. Our multiagent model is presented in the following way. In Section 3.1, we explain the different types of agents in our system, as well as their goals. In Section 3.2, we introduce the system’s organization model, which concerns a micro-macro layered system, and three types of agents' interaction occurring at the micro-level. Besides, according to [13], in order to ensure the emerged solution, it is necessary for a selforganizing multiagent system to find means to control its emergent property. Our control mechanism will be explained through agents' conversation in Section 4.3. 3.1 The Agents By analyzing the system requirements, we associate one manager with one entity. In the structural modeling domain, the entities in question are geological horizons and geological faults. Therefore, we define an agent Horizon Manager for each geological horizon, and an agent Fault Manager for each geological fault. Each agent’s goal is to validate and to update the geological structural interpretations related to its associated entity. These domain-specific tasks are fulfilled with the help of its knowledge base. Additionally, three more agents are defined in the system: 1) Drilling Data Manager, whose goal is to complete the drilling marker information recorded in a trajectory, and then to distribute them to related Horizon Managers or Fault Managers; 2) Earth Model Viewer, whose goal is to present a GES view corresponding to the full earth model’s structural description; 3) Model Checker, whose goal is to detect any global inconsistency in the full earth model's structural description that is represented in GES. 3.2 The System Organization An agent’s expertise and information is related to its embedded entity. Thus it sometimes needs to collaborate with others to accomplish its validation and update tasks. In order to describe the agents’ interaction, we define our system organization in a micro-macro layered model (cf. Fig. 6). The micro-level is composed of Horizon Managers and Fault Managers. For such an agent, each of the local structural interpretations reflects a chronologicaltopological relation from its associated surface to another agent's. We then consider that two agents linked by such a relation are neighbors. For the maintaining purpose, each agent reacts to the modification in its local neighborhood. But due to the restriction of local information, as well as local problem-solving capacity, agents cooperate by exchanging their information. Agent's interaction is then initiated for different kinds of cooperation, which can be classified into three scenarios:
Using Multiagent System to Build Structural Earth Model DDM: Drilling Data Manager EMV: Earth Model Viewer MC: Model Checker HM: Horizon Manager FM: Fault Manager
MC DDM
HM
HM
EMV
Macro-level
FM
HM HM
FM
43
FM
HM
Micro-level
c
Fig. 6. The multiagent system overview in a micro-macro layered organization model
• Delegate Check: A micro-level agent searches to check its current neighborhood, with the partial drilling information distributed from the macro-level. If the agent’s local knowledge is not sufficient to fulfill this task, it delegates the check to its current neighbors. This check delegation continues between the neighboring agents at this level, until either a sure result occurs or no more agents are left to delegate. • Propose Neighbor: When a current neighborhood is checked to be invalid, it has to be updated. In that case, a micro-level agent (playing the role of supplier) proposes a new neighboring relationship to another micro-level agent (playing the role of replier), in regard to the drilling information. Both agents verify if integrating the new relationship into their current neighborhood will give rise to conflicts. In the case that no conflict is detected, the supplier agent waits for a reply (either accept or reject) of the proposal from the replier agent. In case of an acceptance, the neighborhood of the both agents will be updated. In case a conflict is detected, a Handle Neighboring Conflict conversation will be initiated immediately. • Handle Neighboring Conflict: When a conflict is detected in a Propose Neighbor scenario, either in the supplier side or in the replier side, resolving it involves three micro-level agents: the agent detecting the conflict in its current neighborhood (playing the role of conflict-holder); the agent invoking the conflict (playing the role of conflict-maker); and the agent capable to solve the conflict (playing the role of conflict-solver). The conflict-holder and the conflict-maker roles are respectively played by the proposal's supplier agent and replier agent, or vice versa. The conflict-solver role is played by a current neighbor of the conflict-holder. It is with this agent that the neighboring conflict occurs. In order to handle this conflict, the conflict-holder needs to know the chronological-topological relation between the conflict-solver's surface and the conflict-maker's surface. That is out of its knowledge. So it initiates a Handle Neighboring Conflict conversation to send request to the conflict-solver. Once it gets the response, it can handle the conflict by updating its current neighborhood. The macro-level reflects the view concerning the global earth model. It consists of three agents: Drilling Data Manager, Earth Model Viewer, and Model Checker. The Drilling Data Manager is assigned the task of completing and distributing the drilling marker information. So it needs to know all the markers encountered while a drilling progresses inside the full model. The Earth Model Viewer presents the global GES view of the earth model, which is in accordance with all the micro-level agents’ current neighborhood. Since a GES view assembles various surfaces' interpretations as a directed acyclic graph (DAG). The Model Checker is then defined to detect the earth model's global inconsistencies as graph-related problems in a DAG.
44
B. Zhu et al.
4 A Cognitive Agent Model This section presents the cognitive agent model defined for the micro-level agents. This agent model is developed with a knowledge model for accomplishing the domain-specific tasks, a behavior model for pursuing its goal, and a communication model for controlling its collaboration with other agents. 4.1 Structural Modeling Expertise In order to accomplish the domain-specific tasks, we have developed two expertise models [12] for Horizon Manager and Fault Manager respectively. Each one consists of: a structural modeling ontology [14], a set of rule-based inferences [14], and several problem solving methods (PSMs). The ontology is common between the two types of agents. So it is defined out of the agent model, but shared by different agents. The inferences and the PSMs are respectively defined inside Horizon Manager and Fault Manager, so as to distinguish different reasoning processes for interpreting horizons and faults. An agent’s local knowledge is then represented by its expertise model and the existent structural interpretations concerning its embedded surface. In order to help to well understand how such a cognitive agent reasons, step by step, by using the modeled domain-specific inferences to solve a local problem, we give out two examples. The first example consists of two rules used for deciding a chronological neighborship. IF
(surface A is a Fault) ^ (surface B is a Horizon) ^ (surface B is not topographical) ^ (surface A and B are in the same geological block) THEN surface A is older than surface B IF (surface A is a Horizon) ^ (surface B is a Horizon) ^ (surface A is erosional) ^ (surface B is parallel) ^ (surface A and B are in the same geological block) THEN surface A is younger than surface B
The second example is illustrated by the PSM shown in Fig. 7. It concerns a Handle Neighboring Conflict scenario, in which three agents interact to solve a neighboring conflict: agent A1detects conflict in its current neighborhood; A2 is A1's current neighbor, with which A1 is in conflict; agent A3 proposes to A1 a new chronological relation, which then causes A1's conflict. In this situation, A2 is the agent capable to solve the conflict. We present the Conflict handling PSM of A2, who is supposed to be a Horizon Manager in this example. Its objective in this scenario is to provide the conflict-holder A1 with a conflict-resolving solution. In this example, by receiving conflict information from the conflict-holder A1, A2 tries to solve this conflict by finding a chronological-topological relation with the conflict-maker A3. The modeled rules implicating such relations for a horizon are used to execute this behavior. A new relation is then proposed to the conflict-maker A3. By referring to this relation plus the conflict resolution indication rules, A2 chooses a solution for resolving the received conflict. Accordingly, A2 keeps or updates its current neighborhood with the conflict-holder A1, and informs A1 the decided resolution, so that A1 knows how to handle its conflict.
Using Multiagent System to Build Structural Earth Model
45
Fig. 7. The problem solving method that represents the reasoning process of solving a neighboring conflict
For the moment, besides Conflict handling PSM, we have developed Delegate Checking, Conflict Detecting, and Neighborhood Proposing PSMs for our cognitive agent model. Agents use them to fulfill different local tasks in various interaction scenarios. 4.2 Agent's Behavior Model Our agent's behavior specifies the eco-agent's meta-behavior, which has been illustrated by an automaton in [14]. In our system, a cognitive agent's goal consists in ensuring that its local neighborhood, which reflects the geological structure interpretations of its associated surface, is consistent with the marker information recorded along a drilling log. So an agent’s satisfaction is dependent on a coherent local neighborhood. To try to be satisfied, agent analyzes and modifies its local neighborhood. The modification may raise conflict with its current neighbors. Thus, this agent needs to know the chronological-topological relation between the surface of its current neighbor, with which the conflict occurs, and the surface of the conflict-maker. It then asks for the help of its neighbor agent. In order to find out this relation, the neighbor agent uses its structure modeling expertise, by considering the geological characteristics of its own surface and that of the conflict-maker. By referring to the found relation, this neighbor agent replies to the conflict-holder agent with a conflictsolving resolution. Meanwhile, it tries to propose this relation to the conflict-maker, and updates its current neighborhood to integrate the new relation. After receiving the resolution, the agent in conflict tries to update its current neighborhood, so as to reach its satisfaction by making this conflict disappear. 4.3 Agent's Prioritized Conversation Due to the limitation of its local knowledge, an agent needs to communicate with the others which can help it to perform its local tasks. Therefore, several agents are grouped by the message flow circulating in an identified conversation. According to their knowledge, they participate in the conversation with different roles. For example, in a conversation dedicated to Propose Neighbor issue, the agent who proposes a
46
B. Zhu et al.
new neighboring relation plays the role of supplier, the agent who accepts or rejects the proposal plays the role of replier. Even though no conflict is raised in the supplier side, the supplier still does not know whether the proposed neighboring relation can be established or not. In that case, it is the replier who is capable to detect the conflicts in its side, to initiate a conflict-solving conversation in need of resolving, and to decide if the proposed relation can be established or not. Since each agent searches to reach its goal by interacting with others, various conversations simultaneously occur in the system. Therefore an agent needs to know how to select the first conversation in which it should participate with a particular role, so as to decide its behavior in front of different situations. For this reason, we define that each conversation has a priority value, which allows an agent to select an identified conversation with a specified participant role. According to different predefined interaction issues, the priority value of a scenario-typed conversation can be statically distinguished. For example, detecting and solving a neighboring conflict should be executed before validating the current neighborhood with a piece of new drilling marker information. Because it is indispensable to have a stable neighborhood before checking it. Therefore, a Handle Neighboring Conflict conversation has priority over a Delegate Check conversation. In front of different scenario-related conversations, the agent selects one having the highest priority to participate. While various conversations are related to a same scenario type, the agent selects the conversation whose related message is the first one in its mail box. By participating in a conversation, agent plays a particular role, which then allows to decide if it is satisfied or if it needs to escape. That then permits to specify the appropriate behaviors it should have in searching for its satisfaction, and in searching for and in performing an escape. An agent quits a conversation when it becomes satisfied in playing its participant role. Meanwhile, it abandons this role. The agent continues to participate in other conversations until there remains no more conversations to join. At that time, the agent plays the role of validator in its local environment. And it stays satisfied in case no more drilling marker information comes. While all the agents achieve this state, there is no more interaction occurring in the system, the agents' organization then becomes stable. Based on the Contract Net protocol [5], we have defined three interaction protocols for modeling the three conversation scenarios: Delegate Check, Propose Neighbor, and Handle Neighboring Conflict. The parameter timeout helps to eliminate the conversations that have waited a long time for its participant agents’ replies. That permits to avoid having too many interactions at the same time, while the number of Horizon Manager and Fault Manager grows.
5 Experimentation The above described approach has been implemented in DIMA platform [6]. In this section, we present the experimentations obtained by testing two simulated examples. Both are composed of a geological scene pictured in cross-section and of the simulated drilling trajectories passing through the earth model. In order to introduce the drilling information into the system, the markers recorded in the drilling logs are written in the XML format.
Using Multiagent System to Build Structural Earth Model
F2
47
F1
Trajectory
e c
o2 o1
d b o2
a
c c o2
Fig. 8. The left side is the cross-section of a simple geological scene, on which the dotted line represents a 1D sampling trajectory. The right side presents its GES view. F2
E1
l
F1
k e c
o2 o1
d b o2
a
c
a
E2
b
Tra1
c
o1
o2
k
E2 l
Tra2
Fig. 9. A geological scene with two simulated drilling trajectories Tra1, Tra2. The GES in the middle is resulted from Tra1. After validating with Tra2, it evolves to the GES in the right side.
In the first example (cf. Fig. 8), the current GES is empty. By validating with the trajectory, a set of chronological relations are proposed and established. Therefore, a GES corresponding to the model is initialized as shown in the right side. The second example is tested with a more complex geological scene (cf. Fig. 9), which involves two intersecting faults, parallel, erosional and onlap horizons. In this example, we simulate two trajectories Tra1 and Tra2. Tra2 is more complex, because it meets more surfaces. Moreover, it is more difficult to manage the drilling polarity, such as from-recent-to-ancient and from-ancient-to-recent, between two successive markers. By considering Tra1, we have built the model’s first GES view. By validating with Tra2, we have then evolved this initialized GES view (cf. Fig. 9).
6 Conclusion In this paper, in order to tackle the complexity of predicting global strategy for validating and improving a full earth model’s structural description, we proposed to resolve the global problem by favoring the self-organization of cognitive agents, each of which is related to an individual surface in question. Our attention is then focused on acquiring and modeling the adequate knowledge for interpreting each geological
48
B. Zhu et al.
surface, and on controlling these agents' interaction. An emergent solution is achieved thanks to agents’ collective behaviors. This approach makes the system more flexible, and also facilitates the knowledge acquiring and modeling work. Our approach has been validated with the simulated examples defined by the domain experts. The perspective of this approach can be: 1) detecting the correlation errors in seismic data interpretation; 2) merging the GESs corresponding to the different cross-sections of a same geological model, so as to provide a coherent GES view describing the whole earth model's geology.
References 1. Brandel, S., Schneider, S., Perrin, M., et al.: Automatic Building of Structural Geological Models. Journal of Computing & Information Science & Engineering 5(2), 138–148 (2005) 2. Drogoul, A., Ferber, J., Jacopin, E.: Pengi: Applying Eco-Problem-Solving for Behavior Modelling in an Abstract Eco-System. In: Modeling and Simulation: Proceedings of ESM 1991, Simulation Councils, pp. 337–342 (1991) 3. Dury, A., Le Ber, F., Chevrier, V.: A Reactive Approach for Solving Constraint Satisfaction Problems. In: Rao, A.S., Singh, M.P., Müller, J.P. (eds.) ATAL 1998. LNCS, vol. 1555, pp. 397–411. Springer, Heidelberg (1999) 4. Ferber, J.: Multi-Agent Systems: An Introduction to Distributed Artificial Intelligence. Addison-Wesley, Reading (1999) 5. Foundation for intelligent physical agents: FIPA Contract Net Interaction Protocol Specification. FIPA TC Communication, SC00029H, Geneva, Switzerland (1996–2002) 6. Guessoum, Z., Briot, J.P.: From Active Object to Autonomous Agents. IEEE Concurrency 7(3), 68–78 (1999) 7. Jennings, N., Sycara, K., Wooldridge, M.: A Roadmap of Agent Research and Development. Journal of Autonomous Agents and Multi-Agent Systems 1(1), 7–38 (1998) 8. Ohira, T., Sawatari, R., Tokoro, T.: Distributed Interactions with Compution. In: Agents breaking away, Proc. of Modelling Autonomous Agents in Multi-Agent World (1996) 9. Perrin, M.: Geological Consistency: An Opportunity for Safe Surface Assembly and Quick Model Exploration. 3D Modeling of Natural Objects, A Challenge for the 2000’s, 3(pp. 4–5), Nancy France (June 1998) 10. Perrin, M., Zhu, B., Rainaud, J.F., et al.: Knowledge-driven Application for Geological Modelling. Journal of Petroleum Science and Engineering, 89–104 (2005) 11. Ruas, A.: Modèle de généralisation de données géographiques à base de contraintes et d’autonomie. Phd thesis, Université de Marne la Vallée (1999) 12. Schreiber, G., Akkermans, H., Anjewierden, A., et al.: Knowledge Engineering and Management: The CommonKADS Methodology. MIT Press, Cambridge (2000) 13. Serugendo, G., Gleizes, M.-P., Karageorgos, A.: Self-organisation and emergence in MAS: An overview. Informatica 30, 45–54 (2006) 14. Zhu, B., Guessoum, Z., Perrin, M., et al.: Using multiagent systems for validating geological interpretation. In: 2nd IEEE International Conference on Intelligent Computer Communication and Processing, pp. 75–81 (2006)
Agent-Supported Protein Structure Similarity Searching Dariusz Mrozek, Bożena Małysiak, and Wojciech Augustyn Silesian University of Technology, Department of Computer Science, Akademicka 16, 44-100 Gliwice, Poland [email protected], [email protected], [email protected]
Abstract. Searching for similar proteins through the comparison of their spatial structures requires efficient and fully automated methods and become an area of dynamic researches in recent years. We developed an algorithm and set of tools called EAST (Energy Alignment Search Tool). The EAST serves as a tool for finding strong protein structural similarities in a database of protein structures. The similarity searching is performed through the comparison and alignment of protein energy profiles received in the computational process based on the molecular mechanics theory. This representation of protein structures reduces the huge search space. In order to accelerate presented method we implemented it with the use of Multi Agent System (MAS). This significantly improved the efficiency of the search process. In the paper, we present the complexity of the search process, the main idea of the EAST algorithm and brief discussion on the advantages of its implementation as MAS.
Unfortunately, this is a very complicated task because of three reasons: 1. proteins are very complex, usually composed of thousands of atoms; 2. the searching process is usually carried through the comparison of a given structure to all structures in a database; 3. number of protein structures in databases, like Protein Data Bank (PDB) [13] rises exponentially every year and is now 44 578 (July 17, 2007). The first problem is usually solved by decreasing the protein structures complexity in the search process. The most popular methodologies developed so far base on various representations of protein structures in order to reduce the search space. A variety of structure representations were proposed so far, e.g. secondary structure elements (SSE) in VAST [6], locations of the Cα atoms of a protein body and intermolecular distances in DALI [7], aligned fragment pairs (AFPs) in CE [8], or 3D curves in CTSS [9], and many others. These methods are appropriate for homology modeling or function identification. During the analysis of small parts of protein structures that can be active sites in cellular reactions it is required to use more precise methods of comparison and searching. The EAST algorithm [12] that we developed benefit from the dependency between the protein structure and the conformational energy of the structure [10]. In our research, we calculate so called energy profiles (EPs) that are distributions of various potential energies along proteins chains (section 2). This reduces the structure complexity. However, a reasonable searching time is still difficult to achieve in the classic client-server architecture. To solve the problem we decided to reimplement our EAST method to work in the Multi Agent System (MAS). We treat agents as autonomous programs, executed in given place, able to communicate and learn [11]. Since the searching process is carried through the pairwise comparison of given query structure to all structures in a database it is possible to distribute this task to many distinct machines and accelerate entire method. In the paper, we briefly describe the idea of energy profiles (section 2) and our method of similarity searching (section 3), the architecture of the multi-agent system that support the search process (section 4), the environment tests (section 5) and an example of the system usage (section 6).
2 Protein Structure Energy Profiles In our research on active sites of enzymes, we calculate distributions of conformational energy along amino acid chains of proteins. However, amino acids (peptides) are not directly considered in the calculation process. On the contrary, all performed calculations base on Cartesian coordinates of small groups of atoms that constitute each peptide. Therefore, energy distributions can be seen as forms of representation of protein structures. The distribution of different energies along the protein/enzyme polypeptide chain may be very descriptive for protein function, activity and may reflect some distinctive properties. Definition 1: We define a single protein energy profile EP as a set of energy characteristics of various type of energy, determined for a given protein structure.
Agent-Supported Protein Structure Similarity Searching
EP = E st , E ben , E tor , E vdw , E el , E tot
51
(1)
where: Est denotes a distribution of the bond stretching energy, Eben is a distribution of the angle bending component energy, Etor is a distribution of the torsional angle energy, Evdw is a distribution of the van der Waals energy, Eel denotes a distribution of the electrostatic energy. Etot is a distribution of the total energy which is a summary of all component energies in each residue. Definition 2: The energy distribution Et, where t is a type of energy, is called energy characteristics. Energy characteristics are ordered sets of energy values (so called energy points) calculated for groups of atoms of consecutive residues in the polypeptide chain. Let the R be an ordered set of m residues in the polypeptide chain R = {r1r2 r3 ...rm }, and Xin be a set of atomic coordinates building the ith residue (n is a number of atoms of the ith residue depending on the type of residue), then a simplified protein structure can be expressed by the ordered set of small groups of atoms X = {X 1n X 2n X 3n ... X mn }. For the set X we calculate energy characteristics for each type of energy t in the protein energy profile:
{
}
E t = e1t e 2t e3t ...e mt ,
(2)
where: eit is an energy point of the ith residue, t is a type of energy. Energy characteristics (or energy profiles) represent protein structures in a reduced form of ordered sets of energy points, just like other algorithms as sets of positions of Cα atoms. This reduces the search space during the comparison of two protein structures. In our approach, we compute energy characteristics base on the protein atomic coordinates retrieved from the public, macromolecular structure database Protein Data Bank (PDB) [13]. During the calculation we use the TINKER [14] application of molecular mechanics and the Amber [15] force field which is a set of physicalchemical parameters. In this way, we build complete energy profiles as sets of characteristics and store them in our database called Energy Distribution Data Bank (EDB). Afterwards, protein structures can be compared to each other based on their energy profiles, in order to find strong structural similarities, places of discrepancies, or possible mutations.
3 Searching Process with the Use of Energy Profiles We can use energy profiles to search for structurally similar proteins or search just for some particular parts of their structures. The search process is performed on the energy level with the use of our EAST algorithm and profiles stored in our Energy Distribution Data Bank (EDB). The process is realized through the comparison and alignment of energy characteristics of a query protein and each candidate protein from the EDB. In the comparison we consider only one selected energy type from each EPs. The alignment can be thought as the juxtaposition of two energy characteristics that gives the highest number of identical or similar energy points (residues). A
52
D. Mrozek, B. Małysiak, and W. Augustyn
similar word means that energy points do not have to be identical but they should indicate a similarity with the given range of tolerance. This tolerance is different for different types of energy considered in the search process. As a consequence of a suitable juxtaposition and based on appropriate similarity measures it is possible to evaluate a similarity degree of compared molecules and optionally find regions of dissimilarity. Therefore, the EAST algorithm examines identities, similarities (both are qualified as matches) and disagreements (mismatches) of compared energy points of both energy characteristics. Analogically to the nucleotide or amino acid sequences alignment, some mismatching positions and gaps can appear in the best alignment of energy characteristics. These mismatches and gaps reflect possible evolutionary changes in different organisms. During the search process with the use of our EAST algorithm a user specifies a query protein (or a part of a molecule) with a known structure. This query-structure is then transformed into the energy profile (EP). Query-molecule EP (QEP) can be compared with other energy profiles stored in the EDB. This is a pairwise comparison – each pair is constituted by the query energy profile QEP and energy profile of the successive molecule from the EDB (candidate energy profile, CEP). In the pairwise comparison, both QEP and CEP are represented by only one chosen energy characteristic (e.g. torsion angle energy, bond stretching energy, electrostatic, or other). During this phase we build a distance matrix D to check the distance between energy characteristics of the query molecule and all candidate molecules. The distance matrix D allows to compare all energy points of the query molecule energy characteristic E At = (e tA,1 e tA, 2 ...e tA, n ) (part of the QEP) to all energy points of the
(
)
candidate molecule energy characteristic E Bt = e Bt ,1 e Bt , 2 ...e Bt , m (part of the CEP) and is filled according to the expression (3). In the energy distance matrix D, the entry dABij denotes the distance between energy points of the ith residue of protein A and the jth residue of protein B, and can be expressed as: t t , d tAB ,ij = e A ,i − e B , j
(3)
where: t is a type of energy that is considered in the searching. Based on the distance matrix D we perform the optimization of the alignment path. To optimize the alignment path we use modified, energy-adapted Smith-Waterman method (originally published in [17], with later refinements presented in [18], [19], [20]) that produces the similarity matrix S according to the following rules: for 0 ≤ i ≤ n and 0 ≤ j ≤ m : Si 0 = S 0 j = 0 ,
S ij(1) = S i −1, j −1 + ϑ ( d ijAB ) , S ij( 2 ) = max {S i − k , j − ω k } , 1≤ k ≤ n
S
(3) ij
= max {S i , j − l 1≤ l ≤ m ( 4) ij
S
− ωl } ,
= 0,
S ij = max {S ij( v ) } , v =1..4
(4)
Agent-Supported Protein Structure Similarity Searching
53
where: ω k , ω l are gap penalties for horizontal and vertical gaps of length k and l, respectively, and ϑ (d ) is a function which takes a form of similarity award ϑ + ( d ijAB ) for matching energy points ( e tA,i and e Bt , j ) or a form of mismatch penalty ϑ − ( d ijAB ) (usually a constant value) in the case of mismatch (Fig. 1). The match/mismatch (Fig. 1) is resolved based on distance d ijAB between considered points (stored in the distance matrix D) and additional parameter d0 called cutoff value. The cutoff value is the highest possible difference between energy points e tA,i and e Bt , j when we treat these two points as similar. The cutoff value determines the range of tolerance for energy discrepancies. The cutoff value can differ for various types of energy, e.g. for the torsional angle energy the default value was established to d0=1.4 Kcal/mole based on a priori statistics [12]. Default settings for the energy-adapted Smith-Waterman method are: mismatch penalty ϑ − ( d ijAB ) =−1/3, affine gap penalty ωk=1.2+1/3*k [21], where k is a number of gaps and dij is single cell value of the distance matrix D. The similarity award ϑ + ( d ijAB ) for a match depends on distance dij. In the scoring function ϑ (d ) in Fig. 1 we have to define two values: a cutoff value d0, and an identity threshold did. The identity threshold did is the highest possible difference between energy points when we treat two points as the same. The value of the identity threshold was chosen arbitrary to 0.3-0.5 Kcal/mole during observations of energy characteristics performed for many proteins [12]. The calculation of similarity between two energy points based on their distance is presented in Fig. 1. For each distance we calculate a similarity coefficient (similarity award/penalty for a single match/mismatch). These degrees of similarity are aggregated during the execution of the Smith-Waterman method and contribute to the cumulated Smith-Waterman Score. Additional similarity measures are also calculated: RMSD (Root Mean Square Deviation) and Score [12]. The energy-adapted Smith-Waterman method is one of the most important and time consuming parts of the EAST search algorithm. In the Smith-Waterman method cumulated value of similarity score (Smith-Waterman Score) rises when considered energy points match to themselves (are equal or similar with the given range of tolerance d0), and decreases in regions of dissimilarity (mismatch). Moreover, the energyadapted Smith-Waterman algorithm with the similarity award/penalty given by a function (not constant values) considers both, number of matching energy points and a quality of the match in the final alignment [16].
Fig. 1. Similarity award/penalty ϑ (d ) allows to measure the similarity based on the distance between two energy points of compared molecules
54
D. Mrozek, B. Małysiak, and W. Augustyn
4 Architecture of the MAS-EAST Search System In order to accelerate our EAST algorithm we implemented it in the Multi-Agent System. To this purpose we used JADE (Java Agent DEvelopment Framework) which is an open source platform for peer-to-peer agent based applications. The JADE simplifies the implementation of multi-agent systems through a given middleware and through a set of tools that supports the debugging and deployment phases [22]. The most time consuming phase of the EAST algorithm is the optimization (alignment) phase which uses the energy adapted Smith-Waterman method. The method is executed n times for each pair: query molecule vs. candidate molecule, where n is a number of candidate molecules in the EDB. Therefore, the whole alignment task is divided into many subtasks – each one realizes the alignment phase for a smaller group of molecules. In our system we implemented 2 types of agents: 1. Coordinator Agent, CA (only one) – responsible for a division of tasks, sequencing and sharing ranges of molecules to compare, and the consolidation of results. 2. Alignment Agents, AAs (many) – responsible for the main activity that are: a comparison of a query structure with a portion of structures from the EDB, an appropriate alignment, and similarity measures generation. Each AA does the same alignment task for a different set of candidate molecules from the EDB. The MAS-EAST searching process consists of the 3 phases: 1. Initialization – the parameters of the searching process and query energy characteristic are passed to Alignment Agents (Fig. 2). 2. Alignment – each AA performs the alignment for a given range of molecules from the EDB. When an AA finishes its work it returns results to the CA and can get another portion of molecules to work with (Fig. 3). 3. Consolidation – the CA consolidates results of alignments obtained from AAs. The Initialization phase and the Alignment phase are launched sequentially. Nevertheless, the comparison of some portion of structures and consolidation of partial results from the other AAs may occur parallel. The Initialization phase consists of the following steps (Fig. 2): 1. A user executes the search process (through the GUI) that activates the Coordinator Agent (CA). Searching parameters are passed to the CA. 2. The CA requests appropriate energy characteristic of the Query EP from the EDB. 3. The CA obtains appropriate characteristic as a sequence of energy points. 4. The CA distributes the query energy characteristic and searching parameters to all Alignment Agents (AA). The Alignment and Consolidation phases consist of the following steps (Fig. 3): 1. The AAs request a portion of data (k candidate molecular structures represented by energy characteristics, CECs) to compare. This portion is called the package. 2. The retrieved CECs (package) are sent to AAs. 3. For the query molecule and each candidate molecule from obtained package each AA runs a comparison process including: distance matrices, alignment using energy-adapted Smith-Waterman method, and computation of similarity measures.
Agent-Supported Protein Structure Similarity Searching
55
Fig. 2. Communication between agents in the Initialization phase
4. The AAs send results of comparison to the CA. The CA consolidates the partial results asynchronously. 5. The CA sends final results to the user’s GUI window and stores them in the EDB (Except of running the searching process, the results can be simply selected from the database when next time any user executes a query with the same searching parameters). The number of molecules to compare in one package is constant for a searching process, is configurable and is one of the parameters sent in the Initialization phase. Packages are retrieved by AAs continuously until all molecules from the EDB are compared. The CA coordinates the process sending the ranges of molecules to be compared by each AA. The ranges of molecules are disjoint. Therefore, alignment subtasks can be completed independently. If any of the AAs fails its task or disconnects from the system the CA catches the fact and the package of molecules is passed once again to another free AA. There is no need to resolve any conflicts in the situation because results are sent to the CA after the whole package is processed. We considered other architectures of the system and different interactions between agents in the previous implementations of the MAS-EAST algorithm. In the first implementation, all packages were delivered to AAs by the CA. However, this slowed down the retrieval process, which had to be carried out sequentially. In the second implementation, the CA divided the whole alignment task to the current number of connected AAs. As a result, the number of packages given to each AA was balanced. This caused many problems, e.g. slower machines completed their tasks much longer than faster machines, a failure of one AA caused a failure of the whole search process, there was no possibility to rescale the alignment task and join additional AAs during the search process. The current implementation (Fig. 2 and Fig. 3) solves all these problems and allows to take over unfinished tasks in the case of failure. In the future, we consider using more Coordinator Agents to secure the system against crashes of the single CA.
56
D. Mrozek, B. Małysiak, and W. Augustyn
Fig. 3. Communication between agents in the Alignment phase (solid lines) and Consolidation phase (dashed lines)
5 Tests of the MAS-EAST We tested the agent supported EAST method for many molecules using a set of batch tests. In our experiments we used the EDB database containing 2 240 EPs of proteins from our mirror of the Protein Data Bank (containing more than 44 000 proteins). The subset was chosen arbitrary. Tests were performed with the use of many computers with good performance abilities (from Intel Pentium ® 4 CPU 2.8 GHz, 1GB RAM to Intel ® Core 2 CPU 2.13 GHz, 1 GB RAM). Each agent resided on distinct machine. Tests confirmed our expectations – the EAST algorithm implemented in the multi agent system is faster than EAST working in standard client-server architecture. The speedup depends on number of Alignment Agents working together to complete the search process. In Fig. 4 we present average execution times (search times) as a function of the number of working Alignment Agents. Results are presented for three molecules that considerably differ with the number of amino acid (residues) in their chains. In Fig. 4 we can observe the execution time decreases with a growing number of working AAs in the system. At first, the decrease is significant, e.g. for 2 working AAs the search process takes half of the time it takes for 1 working AA. For molecule 1QQW the average time we get results is 34 seconds (3AAs), 56 seconds (2AAs), 109 seconds (1 AA). In the client-server implementation of EAST [12] it took 201 seconds but tests were run on slower machine. Afterward, the decrease is not so significant and finally, for 10 working AAs the search process takes about 16 seconds. The non-linear decrease is probably caused by the communication with the EDB database, data retrieval and delivery. However, this implementation gives us a really good acceleration – for 10 working AAs the average acceleration ratio is 6.33 (7.5 for molecule chains containing 150-450 residues, 3.8 for small molecules up to 100 residues). In the standard client-server implementation of the EAST algorithm (CS-EAST) we could observe the strong dependency between the average execution time and the length of the query protein chain [16]. Longer protein chains cause the necessity to build bigger distance matrices and elongate the alignment process. In Fig. 5 we show the average execution times as a function of chain lengths for the MAS-EAST implementation, a number of working AAs is constant (for a single series of data).
Agent-Supported Protein Structure Similarity Searching 120
time [s ]
57
Average execution times as a function of the number of AAs (number of residues = const.)
Fig. 4. Average search times as a function of the number of working Alignment Agents for 3 molecules from PDB/EDB: 1R3U (Hypoxanthine-Guanine Phosphoribosyltransferase), 3ERK (Extracellular Regulated Kinase 2), 1QQW (Human Erythrocyte Catalase) 120
time [s]
Average execution times as a function of chain lengths (number of AAs = const.)
100
1 AA 2 AAs 4 AAs 6 AAs 10 AAs
80
60
40
20
chain length [res idues]
0 32
80
110
178
256
349
454
674
996
Fig. 5. The dependency between average search times and chain length for different numbers of working Alignment Agents (1, 2, 4, 6, 10 AAs)
In Fig. 5 we can observe the dependency between an execution time and chain length is stronger for the search system with a low number of working AAs. For more AAs working in the system the search time is more independent on chain length. This is an additional advantage of the MAS-EAST implementation. Our tests shown the type of energy used in the search process does not influence the execution time. The package size during tests was set to 20 molecules. This makes the system more scalable and enables to join additional AAs in the middle of the running process.
6 Example of the System Usage We tested our method for more than 100 molecules that are in the area of our scientific research. Beneath we present results and analysis of one of the search process
58
D. Mrozek, B. Małysiak, and W. Augustyn
performed in our tests. The process was run for query molecule representing the whole protein. However, it can be executed just for the smaller parts of protein structures representing active sites, biological functional domains, 3D motifs or any other structural pattern. The presented example concerns proteins from the RAB family that are members of the bigger group of GTPases – particles that have the ability to bind and hydrolyze GTP molecules. Therefore, proteins of the RAB family play an important role in intracellular reactions of living organisms. They are elements of the signal pathways where they serve as molecular controllers in the switch cycle between active form of the GTP molecule and inactive form of the GDP. In the Fig. 6 we can see results of the similarity searching process performed for the 1N6H molecule (Crystal Structure Of Human RAB5A). Results are sorted according to the Score similarity measure. Results can be interpreted based on similarity measures: Score and SmithWaterman Score (SW-Score) – the higher value the higher similarity, RMSD – the lower value the better quality of the alignment, and based on output parameters: Length – alignment frame, Matches – number of aligned energy points, Match% – percentage of matching positions in the alignment frame (the higher value the better alignment). Molecules above the horizontal line (Fig. 6) were qualified as similar. The line was inserted manually after the searching process and verification based on similarity measures, annotations of molecules and literature. The molecule 1HUQ in the group of similar molecules has the same function in mouse organism (mus musculus) as query molecule 1N6H in human. The alignment of query molecule 1N6H and candidate molecule 1HUQ on the energy level is presented in Fig. 7. Energy characteristics for both molecules have many matching Best results for job: 2007-07-04 13:13:55 Cut-off: 6.1; id threshold: 0.5; energy type: Charge-charge S-W type: Fuzzy; mismatch: -0.3334; gap open: 1.2; gap ext.: 0.3334 PDBID Chain Length Matches Match% RMSD Score S-W Score ----- ----- ------ ------- ------ ------- -------- --------1N6R A 159 154 96 2.091 73.64 132.84 1N6L A 162 155 95 2.247 68.98 134.78 1N6O A 128 126 98 1.914 65.85 98.01 1N6K A 161 155 96 2.365 65.53 130.45 1N6I A 154 146 94 2.391 61.06 121.69 1HUQ A 146 140 95 2.295 61.01 114.89 1N6P A 136 129 94 2.269 56.85 104.47 1N6N A 146 137 93 2.488 55.06 110.64 1TU3 A 142 138 97 2.763 49.95 96.85 --------------------------1GRN A 28 27 96 3.259 8.29 8.96 1GUA A 38 32 84 3.960 8.08 4.87 1R8Q A 26 25 96 3.357 7.45 5.58 ...
Fig. 6. Results of the similarity searching process with the EAST algorithm (screenshot from the EAST) for molecule 1N6H (Crystal Structure Of Human RAB5A). Parameters – cutoff value: 6.1 Kcal/mole, identity threshold: 0.5 Kcal/mole, energy type: electrostatic.
Agent-Supported Protein Structure Similarity Searching
59
Energy Alignment
Energy [Kcal/mole]
50 0 -50 -100 -150 -200 70
80
90
100
110 Residue number A_1HUQ A_1N6H
120
130
140
Fig. 7. Alignment of energy characteristics for query molecule 1N6H (Crystal Structure of Human RAB5A) and resultant molecule 1HUQ (Crystal Structure of the Mouse Monomeric GTPase RAB5C). Visible parts: residues 68-145. Grey color (bright line) indicates mismatching parts, no gaps visible.
a)
b)
Fig. 8. Comparison of the 3D structures (ribbon representation) of molecules: a) query 1N6H (Crystal Structure of Human RAB5A), b) resultant 1HUQ (Crystal Structure of the Mouse Monomeric GTPase RAB5C). Visualization made with the use of the RasMol [23].
energy points indicating structural similarity. Some parts of these characteristics cover each other what verifies a good quality of the alignment. There are also some parts indicating small conformational differences in the compared structures (e.g. residues 104-110). The comparison of the 3D structures of query molecule 1N6H and resultant molecule 1HUQ presented in Fig. 8 confirms the similarity of these two molecules.
7 Concluding Remarks Similarity searching is one of the most frequent tasks performed in bioinformatics database systems. We developed the EAST algorithm of searching for structurally similar proteins (or their parts) with the use of energy profiles. The energy profiles represent protein structures. This reduces the complexity of the search process and is the first acceleration step. The MAS implementation of the EAST algorithm (MAS-EAST) causes additional significant acceleration. Moreover, the MAS
60
D. Mrozek, B. Małysiak, and W. Augustyn
implementation guarantees scalability, failure protection, and an automatic balance of the work according to the workstation possibilities. This gives an ability to search great volumes of protein structural data that we expect to have in the nearest future as an effect of mass Nuclear Magnetic Resonance spectrometry. In the future, we plan to test the MAS-EAST for the EDB containing energy profiles for all molecules in the PDB database and distribute the search problem solution to more computers.
References 1. Fersht, A.: Enzyme Structure and Mechanism, 2nd edn. W.H. Freeman & Co., NY (1985) 2. Dickerson, R.E., Geis, I.: The Structure and Action of Proteins, 2nd edn. Benjamin/Cummings, Redwood City (1981) 3. Lodish, H., Berk, A., Zipursky, S.L., et al.: Molecular Cell Biology, 4th edn. W. H. Freeman and Company, NY (2001) 4. Branden, C., Tooze, J.: Introduction to Protein Structure. Garland (1991) 5. Creighton, T.E.: Proteins: Structures and molecular properties. Freeman, San Fransico (1993) 6. Gibrat, J.F., Madej, T., Bryant, S.H.: Surprising similarities in structure comparison. Curr. Opin. Struct. Biol. 6(3), 377–385 (1996) 7. Holm, L., Sander, C.: Protein structure comparison by alignment of distance matrices. J. Mol. Biol. 233(1), 123–138 (1993) 8. Shindyalov, I.N., Bourne, P.E.: Protein structure alignment by incremental combinatorial extension (CE) of the optimal path. Protein Engineering 11(9), 739–747 (1998) 9. Can, T., Wang, Y.F.: CTSS: A Robust and Efficient Method for Protein Structure Alignment Based on Local Geometrical and Biological Features. In: Proceedings of the 2003 IEEE Bioinformatics Conference (CSB 2003), pp. 169–179 (2003) 10. Burkert, U., Allinger, N.L.: Molecular Mechanics. American Chemical Society, Washington (1980) 11. Wooldridge, M.: Introduction to MultiAgent Systems. Wiley &Sons, Chichester (2002) 12. Mrozek, D., Małysiak, B., Kozielski, S.: EAST: Energy Alignment Search Tool. In: Wang, L., Jiao, L., Shi, G., Li, X., Liu, J. (eds.) FSKD 2006. LNCS, vol. 4223, pp. 696–705. Springer, Heidelberg (2006) 13. Berman, H.M., Westbrook, J., Feng, Z., Gilliland, G., Bhat, T.N., Weissig, H., et al.: The Protein Data Bank. Nucleic Acids Res. 28, 235–242 (2000) 14. Ponder, J.: TINKER – Software Tools for Molecular Design, Dept. of Biochemistry & Molecular Biophysics, Washington University, School of Medicine, St. Louis (June 2001) 15. Cornell, W.D., Cieplak, P., et al.: A Second Generation Force Field for the Simulation of Proteins, Nucleic Acids, and Organic Molecules. J. Am. Chem. Soc. 117, 5179–5197 (1995) 16. Mrozek, D., Małysiak, B., Kozielski, S.: An Optimal Alignment of Proteins Energy Characteristics with Crisp and Fuzzy Similarity Awards. In: Proc. of the IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), pp. 1508–1513 (2007) 17. Smith, T.F., Waterman, M.S.: Identification of common molecular Subsequences. J. Mol. Biol. 147, 195–197 (1981) 18. Gotoh, O.: An Improved Algorithm for Matching Biological Sequences. J. Mol. Biol. 162, 705–708 (1982)
Agent-Supported Protein Structure Similarity Searching
61
19. Sellers, P.H.: Pattern recognition in genetic sequences by mismatch density. Bull. Math. Biol. 46, 510–514 (1984) 20. Altschul, S.F., Erickson, B.W.: Locally optimal subalignments using nonlinear similarity functions. Bull. Math. Biol. 48, 633–660 (1986) 21. Altschul, S.F., Erickson, B.W.: Optimal sequence alignment using affine gap costs. Bull. Math. Biol. 48(5-6), 603–616 (1986) 22. Bellifemine, F., Caire, G., Poggi, A., Rimassa, G.: JADE, A White Paper (2003), http://jade.tilab.com/papers/2003/WhitePaperJADEEXP.pdf 23. Sayle, R., Milner-White, E.J.: RasMol: Biomolecular graphics for all. Trends in Biochemical Sciences (TIBS) 20(9), 374 (1995)
Merging Roles in Coordination and in Agent Deliberation Guido Boella1 , Valerio Genovese1, Roberto Grenna1, and Leendert van der Torre2 1
Abstract. In this paper we generalize and merge two models of roles used in multiagent systems which address complementary aspects: enacting roles and communication among roles in an organization or institution. We do this by proposing a metamodel of roles and specializing the metamodel to two existing models. We show how the two approaches can be integrated since they deal with complementary aspects: Boella [1] focuses on roles as a way to specify interactions among agents, and, thus, it emphasizes the public character of roles. [2] focuses instead on how roles are played, and thus it emphasizes the private aspects of roles: how the beliefs and goals of the roles become the beliefs and goals of the agents. The former approach focuses on the dynamics of roles in function of the communication process. The latter focuses on agents internal dynamics when they start playing a role or shift the role they are currently playing. Keywords: Roles, Organizations, O.O. Modeling, MAS, Security.
1 Introduction In the last years, the usefulness of roles in designing agent organizations has been widely acknowledged. Many different models have been designed. Some of them use roles only in the design phases of a MAS [3], while other ones consider roles as first class entities which exist also during the runtime of the system [4].There are approaches that underline how roles are played by agents [2], other ones on how roles are used in communication among agents in organizations [1]. This heterogeneity of the way roles are dfined and used in MAS risks to be a danger for the interoperability of agents in open systems, since each agent entering a MAS can have a radically different notion of role. Thus, the newly entered agents cannot be governed by means of organizations regulating the MAS. Imposing to all agent designers a single notion of role is a strategy that cannot have success. Rather, it would be helpful to design both multiagent infrastructures that are able to deal with different notions of roles, and to have agents which are able to adapt to open systems which use different notions of roles in organizations. This alternative strategy can be costly if it is not possible to have a general model of role that is compatible, or can be made compatible with other existing concepts. In this paper we generalize and merge two models of roles used in multiagent systems, in order to promote the interoperability of systems. The research question is: How A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 62–73, 2009. c Springer-Verlag Berlin Heidelberg 2009
Merging Roles in Coordination and in Agent Deliberation
63
to combine the model of role enactment by [2] with the model of communication among organizational roles of [1]? We answer these questions by extending to agents a metamodel of roles developed for object oriented systems [5]. The relevant questions, in this case, are: how to introduce beliefs, goals and other mental attitudes in objects, and how to pass from the method invocation paradigm to the message passing paradigm. Then we specialize the metamodel to model two existing approaches and we show how they can be integrated in the metamodel since they deal with complementary aspects. We choose to model the proposals of [1] and [2] since they are representative of two main traditions. The first tradition is using roles to model the interaction among agents in organizations, and the second one is about role enactment, i.e., to study how agents have to behave when they play a role. From one side, organizational models are motivated by the fact that agents playing roles may change, for example a secretary may be replaced by another one if she is ill. Therefore, these models define interaction in terms of roles rather than agents. In Boella [1] roles model the public image that agents build during the interaction with other agents; such image represents the behavior agents are publicly committed to. However, this model leaves unspecified, how given a role, its player will behave. This is a general problem of organizational models which neglect that when, for example, a secretary falls ill, there are usually some problems with ongoing issues (the new secretary does not know precisely the thing to be done, arrangements already made etc.). So having a model of enacting and deacting agents surely leads to some new challenges, which could not be discussed, simulated or formally analyzed without this model. In contrast, the organizational view focuses on the dynamics of roles in function of the communication process: roles evolve according to the speech acts of the interactants, e.g. the commitment made by a speaker or the commands made by other agents playing roles which are empowered to give orders. In this model roles are modeled as sets of beliefs and goals which are the description of the expected behavior of the agent. Roles are not isolated, but belongs to institutions, where constitutive rules specify how roles change according to the moves played in the interactions by the agents enacting them. Dastani [2] focuses, instead, on how roles are played by an agent, and, thus, on the private aspects of roles. Given a role described in terms of beliefs, goals, and other components, like plans, the model describes how these mental attitudes become the beliefs and goals of the agents. In this approach roles are xed descriptions, so they do not have a dynamics like in the model of [1]. Moreover, when roles are considered inside organizations new problems for role enactment emerge: for example, how to coordinate with the other agents knowing what they are expected to do in their role, and how to use the powers which are put at disposal of the player of the role in the organization. The same role definition should lead to different behaviors when the role is played in different organizations. In contrast, it species the internal dynamics of the agents when they start playing (or enacting in their terminology) a role or shift the role they are currently playing (called the activated role). So they model role enacting agents: agents that know which roles they play, the definitions of those roles, and which autonomously adapt their mental states to play the roles.
64
G. Boella et al.
Despite the apparent differences, the two approaches are compatible since they both attributes beliefs and goals to roles. So we study by means of the metamodel how they can be combined to have a more comprehensive model of roles. The paper is structured as follows. In Section 2 we describe the requirements on agents and roles in order to build a metamodel; in Section 3 we formally define the metamodel for roles together with its dynamics; in Section 4 we define the basic notions to model agents that play roles; Section 5 deals with the modeling of enacting agents as in Dastani [2]; Section 6 introduces and models roles to deal with coordination in organizations; in Section 7 we merge Dastani [2] and Boella [1] into the framework introduced in Section 3; Conclusions end the paper.
2 Agents and Roles Since the aim of this paper is to build a metamodel to promote interoperability, we make minimal assumptions on agents and roles. The starting point of our proposal is a role metamodel for object orientation. The relation of objects and agents is not clear, and to pass from object to agents we take inspiration from the Jade model [6]. Agents, differently than objects, do not have methods that can be invoked starting from a reference to the object. Rather, they have an identity and they interact via messages. Messages are delivered by the MAS infrastructure, so that agents can be located in different platforms. The messages are modeled via the usual send-receive protocol. We abstract in the metamodel from the details of the communication infrastructure (whether it uses message buffers, etc.). Agents have beliefs and goals. Goals are modeled as methods which can be executed only by the agent itself when it decides to achieve the goal. As said above, we propose a very simple model of agents to avoid controversial issues. When we pass to roles, however, controversial is- sues cannot be avoided. The requirements to cope with both models of roles we want to integrate are: (a) item Roles are instances, associated in some way to their players; (b) roles are described (at least) in terms of beliefs and goals; (c) roles change over time; (e) roles belong to institutions, where the interaction among roles is specified; (f) the interaction among roles specfies how the state of roles changes over time. In Boella [1] roles are used to model interactions, so agents exchange messages according to some protocol passing via their roles. This means that the agent have to act on the roles, e.g., to specify which is the move the role has to play in certain moment. Moreover, roles interact with each other. Dastanis [2] model specifies how the state of the agent changes in function of the beliefs and goals of the roles it plays. However, it does not consider the possibility that the state of the role change and, thus, it ignores how the agent becomes aware of the changes of beliefs and goals of the role. To combine the two models we have to specify how the interaction between an agent and its role happens when the agent changes the state of the role or the state of the role is changed by some event. A role could be considered as an object, and its player could invoke a method of the role. However, this scenario is not possible, since the roles are strictly related to the institution they belong to, and we cannot assume that the institution and all the agents playing roles in the institution are located on the same
Merging Roles in Coordination and in Agent Deliberation
65
agent platform. So method invocation is not possible unless some sophisticated remote method invocation infrastructure is used. Moreover, the role has to communicate with its player when its beliefs and goals are updated. Given that the agent is not an object, the only way is that a role sends a message to its player. As a consequence, we decide to model the interaction between the agent and the role by means of messages too. Finally, we have to model the interaction among roles. Since all roles of an institution belong to the same agent platform, they do not necessarily have to communicate via messages. To simplify the interaction, we model communication among roles by means of method invocation. The fact that roles belong to an institution has another consequence. According to the powerJava [7] model of roles in object oriented programming languages, roles, seen as objects, belong to the same namespace of the institution. This means that each role can access the state of the institution and of the sibling roles. This allows to see roles as a way to specify coordination. In a sense, roles are seen both as objects, from the internal point of view of the institution they belong to, and as agents, from the point of view of their players, with beliefs and goals, but not autonomous. Their behavior is simply to: (1) receive the messages of their players, (2) execute the requests of their player of performing the interaction moves according to the protocol allowed by the institution in that role, (3) send a message to their players when the interaction move performed by the role itself or by some other role results in a change of state of the role.
3 A Logical Model for Roles In Genovese [5] the model is structured in three main levels: universal, individual and dynamic; here we decide not to talk about the universal level and concentrate ourself on agents dynamics. We dene the formalism of the framework in a way as much general as possible, this gives us an unconstrained model where special constraints are added later. 3.1 Individual Level This level is composed by a snapshot model that describes in a particular moment the relationships between individual players contexts and roles, and a dynamic model which links snapshots and actions modeling how the system changes when an action is executed. In the formalization of the model we use objects as basic elements upon which the model is based. Definition 1. A snapshot model is a tuple < O, R types, I contexts, I players, I roles, Val, I contraints IRoles , I Attributes, I Operations, IAttr > – O is a domain of objects, for each object o is possible to refer to its attributes and operations through πI Attr (o) and πI Op (o), respectively. – R types is a set of types of roles. – I contexts ⊆ O is a set of institutions (referred as institutions).
66
G. Boella et al.
– – – – – –
I players ⊆ O is a set of actors (referred as actors). I roles ⊂ O is a set of roles instances (referred as roles instances). I Attributes is the set of attributes. I Operations is the set of operations. Val is a set of values. I constraints is a set of integrity rules that constraint elements in the snapshot.
The snapshot model has also a few functions: – IRoles is a role assignment function that assigns to each role R a relation on I context x I players x I roles. – IAttr is an assignment function which it takes as arguments an object d ∈ O, and an attribute p ∈ πI Attr (d), if p has a value v ∈ Val it returns it, 0/ otherwise. Generally, when a role instance x is an individual of the type D, we write x :: D. If a ∈ πI Attr (x) we write x.a ∈ I Attributes as the attribute instance assigned to object x, the same holds for elements in I Operations. (i,a,o)∈ IRoles (R) means: “the object o represents agent a playing the role R in institution i”, often writen R(i,a,o), and o is the role instance. 3.2 The Dynamic Model The dynamic model relies on the individual level and denes a structure to properly describe how the framework evolves as a consequence of executing an action on a snapshot. In Section 4 and 5, we describe how this model constraints agentsdynamics. Definition 2. A dynamic model is a tuple < S, TM, Actions, Requirements, D constraints, IActions , IRolest πReq , IRequirementst > – S is a set of snapshots. – TM ⊆ S x IN: it is a time assignment relationship, such that each snapshot has an associated unique time t. – Actions is a set of actions. – Requirements is a set of requirements for playing roles in the dynamic model. – D contraints is a set of integrity rules that constraints the dynamic model. – IActions maps each action from Actions to a relation on a set of snapshots P. IActions (s, a,t) tells us which snapshots are the result of executing action a at time t from a certain snapshop. 1 This function returns a couple in TM that binds the resulting snapshot with time t + 1. In general, to express that at time t is carried action a we write at . – About IRolest , Rt (i, a, o) is true if there exists, at a time t, the role instance R(i, a, o). – πReq (t, R) returns a subset of Requirements present at a given time t for the role of type R, which are the requirements that must be fulfilled in order to play it. 1
Notice that given an action, we can have several snapshots because we model actions with modal logic in which, from a world it is possible to go to more than one other possible world. This property is often formalized through the accessibility relationship. Thus, each snapshot can be seen as a possible world in modal logic.
Merging Roles in Coordination and in Agent Deliberation
67
– IRequirementst is a function that, given (i,a,R,t) returns True if the actor a fills the requirement in πReq (t, R) to play the role R in the institution i, False otherwise. We often write Reqt (i, a, R). Intuitively, the snapshots in S represent the state of a system at a certain time. Looking at IActions is possible to identify the course of actions as an ordered sequence of actions such that a1 ; b2 ; c3 represents a system that evolves due to the execution of a, b and c at consecutive times. We refer to a particular snapshot using the time t as a reference, so that for instance πI Attrt refers to πI Attr in the snapshot associated with t in TM. Actions are described using dynamic modal logic [8], in paricular they are modelled through precondition laws and action laws of the following form: 2(A ∧ B ∧C ⊃ d )
2(A ∧ B ∧C ⊃ [d]E)
(1) (2)
Where the 2 operator expresses that the quantified formulas holds in all the possible words. Precondition law (1) specifies the conditions A,B and C that make an atomic action d executable in a state. (2) is an action law2 which states that if preconditions
A ,B and C to action d holds, after the execution of d also E holds. In addition we introduce complex actions which specify complex behaviors by means of procedure definitions, built upon other actions. Formally a complex action has the following form: p0 ϕ ⊂ p1 ; p2 ....; pm ϕ ; p0 is a procedure name, “;” is the sequencing operator of dynamic logic, and the pi ’s, i ∈ [1, m], are procedure names, atomic actions, or test actions3 . Now we show some examples of actions that can be introduced in the dynamic model in order to specialize the model. Role addition and deletion. For role addition and deletion actions we use, respectively R, i →t a, and R, i ←t a. Then using the notation of dynamic logic introduced above, we write: 2(Reqt (i, a, R) ⊃ R, i →t a ) to express that, if actor a fills the requirements at time t (Reqt (i, a, R) is True), a can execute the role addition action that let him play role of type R. The above definition gives us the possibility to model that a role assignment introduces a role instance: 2( ⊃ [R, i →t a]∃xRt+1 (i, a, x)) or the fact that if a does not already play the role R within institution i, then the role assignment introduces exactly one role instance: 2(¬∃xR(i, a, x) ⊃ [R, i →t a]∃!xRt+1 (i, a, x)). Methods. There are other actions through which is possible to change the model as well, for instance agents may assign new values to their attributes [5]. Here, we will focus on the case in which the attribute’s values can be changed by the objects themselves. What we will do is to define methods of objects with which they can change attributes of their own or those of others. Actually, to simplify the model, we define one single primitive action: sett (o1 , o2 , attr, v), which means that object o1 2 3
Sometimes action laws are called effect rules because E can be considered the effect of the execution of d. Test actions are of the form ψ ?ϕ ≡ ψ ∧ ϕ .
68
G. Boella et al.
sets the value of attr on object o2 to v at time t. If o1 and o2 are autonomous agents, the set(o1 , o2 , attr, v) can be executed only when o1 = o2 . Now, we will of course have that: 2( ⊃ [sett (o1 , o2 , attr, v)]attrt+1 (o2 ) = v) which means that in any state, after the execution of set, if the action of setting this attribute succeeds, then the relevant object will indeed have this value for that attribute. Operations. Elements of our framework come with operations that can be executed at the individual level in order to change the model dynamically, the semantics of each operations can be given exploiting the actions defined for the dynamic model. Suppose, for instance, to have an object individual x :: Person with x.mail address attribute, and an operation x.change mail that changes the value of x.mail address to its argument. Using the set primitive is possible to define how the model evolves after the execution of x.change mail operation trough the following axiom: [x.change mailt (s)]ϕ ≡ [sett (x, x, mail adress, s)]ϕ , where x.change mailt (s) identifies the action carried by x at time t to execute the instance operation x.change mail; objects can execute only operations that are assigned to them by I OS relation. In Section 5 we define exec of certain operations as complex actions because we have to describe a more complex semantics.
4 Enact and Deact Roles In Dastani [2], the problem of formally defining the dynamics of roles, is tackled identifying the actions that can be done in a open system such that agents can enter and leave. Here, four operations to deal with role dynamics are defined: enact and deact, which mean that an agent starts and finishes to occupy (play) a role in a system, and activate and deactivate, which means that an agent starts executing actions (operations) belonging to the role and suspends the execution of the actions. Although is possible to have an agent with multiple roles enacted simultaneously, only one role can be active at the same time. Before diving into modeling the four basic operations to deal with roles, we need to match our framework with a few concepts defined in [2], following we report a list of elements together with their definition and then how they fit in our meta-model: – Multiagent system: In [2] roles are taken into account at the implementation level of open MAS, they belong to the system which can be entered or left by agents dynamically. In our framework is possible to view a system as a context to which are linked all roles that can be played by the agents. – Agent role: A role is a tuple σ , γ , ω . Where σ are beliefs, γ goals and ω rules representing conditional norms and obligations. This definition specifies a role “in terms of the information that becomes available to agents when they enact the role, the objectives or responsibilities that the enacting agent should achieve or satisfy, and normative rules which can for example be used to handle these objectives” [2]. With this view we define, for roles of our framework, a set of complex attributes {beliefs, goals, plans, rules} ∈ I Attr togheter with the operations that represent actions that an agent can carry out when it activates the roles instance choosing it from the set of roles it is playing. – Agent type: We consider an agent type “as a set of agent roles with certain constraints and assume that an agent of a certain type decides itself to enact or deact
Merging Roles in Coordination and in Agent Deliberation
69
a role”. To talk about agent types we use classes introduced in the framework as a specification of agent instances at the individual level, with this in mind we use the PL relationship to link agent classes to agent roles (role’s classes) so that the set of roles that an agent can enact (play), is constrainted by I PL. – Role enacting agent: “We assume that role enacting agents have their own mental attitudes consisting of beliefs, goals, plans, and rules that may specify their conditional mental attitudes as well as how to modify their mental attitudes. Therefore, role enacting agents have distinct objectives and rules associated to the active role it is enacting, and sets of distinct objectives and rules adopted from enacted but inactive roles”. In our framework we define a role enacting agent as a instance x having a set of attributes A that represent the internal structures used to deliberate. A = {beliefsa , objectivesa , plansa , rulesa , enacted roles[], active role} ∈ πI Attr (x) The enacted roles attribute is a role ordered record where each entry with index i corresponds to a triple σi , γi , ωi which represents the set of beliefs, objectives, plans and rules associated to roles instance i enacted by x. As introduced above, the model in [2] identifies four operations to deal with role dynamics, in order to to grasp the fundamental ideas proposed in the cited paper, we redefine the enact, deact, activate and deactivate operations respecting their original meaning. Given an agent x, a role instance i :: R played by x in context c s.t., {beliefsr , objectivesr , plansr , rulesr } ∈ πI Attr (i) {beliefsa , objectivesa , plansa , rulesa , enacted roles[], active role} ∈SA πI Attr (x) {enact, deact, activate, deactivate} ∈ πI Op (x) Next we report the semantics of each operation exploiting the set primitive: x.enactt (i)ϕ ⊂ R, s → x; sett (x, x, beliefsa , beliefsa ∪ beliefsr ); sett (x, x, enacted roles[i], < objectivesr , plansr , rulesr >)ϕ
(3)
x.deactt (i)ϕ ⊂ R, s ← x; sett (x, x, enacted roles[i], null)ϕ
(4)
x.activatet (i)ϕ ⊂ sett (x, x, active role, enacted roles[i])ϕ
(5)
x.deactivatet (i)ϕ ⊂ sett (x, x, active role, null)ϕ
(6)
5 The Public Dimension of Roles In Boella-Van der Torre [9] roles are introduced inside institutions to model the interaction among agents. In Boella [1] the model is specifically used to provide a semantics for agent communication languages in terms of public mental attitudes attributed to roles. The basic ideas of the model are: (1) roles are instances with associated beliefs and goals attributed to them. These mental attitudes are public. (2) The public beliefs and goals attributed to roles are changed by speech acts executed either by the role or by
70
G. Boella et al.
other roles. The former case accounts for the addition of preconditions and of the intention to achieve the rational effect of a speech act, the latter one for the case of commands or other speech acts presupposing a hierarchy of authority among roles. (3) The agents execute speech acts via their roles. In order to maintain the model simple enough, we model message passing extending the dynamic model with two actions (methods) send(x,y,sp) and receive(y,x,sp). Where send(x,y,sp) should be read as the action carried by x of sending a speech act (sp) to y and receive(y,x,sp) is the complementary action of y receiving the message from x. It must be underlined that arguments x and y can be agents or roles. A role only listens for the messages sent by the agents playing it: listen(r)ϕ ⊂ P; played by(r, x)?; receive(r, x, sp); Dϕ These rules define a pattern of protocol where P and D have to be read as possible other actions that can be executed before and after the receive. The reception of a message from the agent has the effect of changing the state of other roles. For example, a command given via a role amounts to the creation of a goal on the receiver if the sender has authority (within the system) over it.
2(authoritysys(r, request ⊃ [receive(r, x, request(r, r , act)))]Grt (act))4 To produce a speech act, the agent has to send a message to the role specifying the illocutive force, the receiver and the content of the speech act: communicate(a)ϕ ⊂ P; send(x, r, sp); Dϕ
6 The Combined Model The two models presented above model complementary aspects of roles: the public character of roles in communication and how agents privately adapt their mental attitudes to the roles they play. In this section we try to merge the two approaches using the metamodel we presented. On the one hand, Boella’s model [1] is extended from the public side to the private side, by using Dastani [2] as a model of role enacting. In this way, the expectations described by the roles resulting from the interaction among agents can become a behavior of agents and they do not remain only a description. On the other hand, Dastanis model [2] is made more dynamic. In the original model the role is given as a fixed structure. The goals of agent can evolve according to the goal generation rules contained in it, but the beliefs and goals described by the role cannot change. This is unrealistic, since during the activity of the agent enacting its role, it is possible that further information are put at disposal of the role and that new responsibilities are assigned, etc. In order to merge the two models within the same framework, we need to add (complex) actions which are able to grasp the dynamics introduced in [1] and [2]. Interactions among agents is done through message passing and, in particular, through actions send and receive introduced in section 6. Next we are going to introduce all the speech-acts and complex actions which are needed to grasp the combined model and then we introduce a running example to clarify their use defining a course of actions in the dynamic 4
request(r, r , act) means: role r asks to r ’s player to do act. authoritysys (r, request) expresses that role r has the authority to make a request within system sys.
Merging Roles in Coordination and in Agent Deliberation
71
model defined in section 3.2. An agent who wants to play a role within an open system has to ask to the system for a role instance; this process is handled by two speech act: ask to play(R) and accept to play(r,A), where the first one is sent from the agent to the system in order to ask to play a role of type R, whereas the second is sent from the system to the agent, together with the identifier of the role instance r and a set A of other role instances present in the system, in order to inform the agent with which roles is possible to interact. Next we report the two effect rules associated: 2( ⊃ [receive(s, x, ask to play(R); send(s, x, accept to play(r, A)] played bysys (r, x, s)
(7)
2( ⊃ [send(x, s, ask to play(R); receive(x, s, accept to play(r, A)] played byag (r, x, s))
(8)
Where s is the system, x the agent, and r a role instance of type R. In this section we use
x,y,z. . . to denote agents, s for the system and r, r , r . . . for role instances. Notice that played bysys (r, x, s) and played byag (r, x, s) refer to two different infrastructures; in Rule 7 is the system that, after having acknowledged the agent request, knows that x is going to play r, whereas in Rule 8 is the agent that becomes aware of the play relation between x and r. To link the two predicates with the logical model introduced in Section 3 we have that: played bysys (r, x, s) ∧ played byag (r, x, s) → R(s, x, r). When we are dealing with a single system we can omit s writing played bysys (r, x) and played byag (r, x). To enact a role, an agent, provided the identifier of the role instance it wants to enact, has to send a message to the role and to wait till the role replies with the information about the state of the role: its beliefs, goal, plans, etc. When the state is received, the agent can enact the role in the same way described by Rule 3 in Section 5. In order to model such interaction we introduce two complex actions tell enact, accept enact and two speech acts accept enact and inform enact. Following the specification of the complex actions: tell enact(x, r)ϕ ⊂ played byag (r, x)?; (send(a1, r1, enact(x, r))ϕ accept enactment(r, x)ϕ ⊂ receive(r, x, enact(x, r)); played bysys (r, x)?; send(r, x, inform enact(< beliefsr , objectivesr , plansr , rulesr >))ϕ
(9)
(10)
When the agent receives the specification of the role he wishes to enact, it can internalize them as in Rule 3: 2( ⊃ [receive(x, r, inform enact(< beliefsr , objectivesr , plansr , rulesr >))] Bx (beliefsr ) ∧ x.enacted roles[r] =< objectivesr , plansr , rulesr >)5
(11)
In this combined view is possible that role’s specifications change dynamically, in that case it is up to the role to send a message to its player each time its state is updated: udpate state(r, x)ϕ ⊂ played bysys (r, x)?; (¬Grt (q) ∧ Grt+1 (q))?; send(r, x, inform goal(q))ϕ
(12)
72
G. Boella et al.
2 system_C 1
agent_A
3
r1::R1 4
8
5 6 9
agent_B
7
r2
Fig. 1. Roles in MAS
Last but not least, we need to model the deactment of a role respecting the formalization as in Rule 4, therefore we introduce two speech acts deact, ok deact and a complex action confirm deact defined as follows: confirm deact(r, x)ϕ ⊂ receive(r, x, deact); played bysys (r, x)?; send(r, x, ok deact)ϕ
(13)
After sending ok deact, the system will not consider anymore agent x as player of r: 2( ⊃ [confirm deact(r, x)]¬played bysys (r, x)
(14)
If it is possible for the agent to deact the role, it will receive an ok deact from its role: 2( ⊃ [receive(x, r, ok deact)]x.enacted roles[r] = null ∧ ¬played byag (r, x))
(15)
Fig. 1 depicts two agents which interact through roles in an open system. At time t the system has already agent B that enacts role r2 as represented by the black arrow which goes from agent B to r2. Following the course of actions that describe how the system evolves: 1. At time t+1 agent A asks to institution system C to play a role of type R1: sendt+1 (agent A, system C, ask to play(R1)) 2. At time t+2 system C replies to agent A assigning to him the role instance r1: sendt+2 (system C, agent A, accept to play(r1, {r2})) 3. At time t+3 agent A wants to enact (internalize) role r1: tell enactt+3 (agent A, r1) 4. At time t+4 role r1 receives the speech act from agent A asking for enactment and accepts it, replying to agent A with its specifications: accept enactmentt+4 (r1, agent A) 5. Once that agent A has enacted the role as in Rule 3 it decides, at time t+5, to activate it and then to ask to the agent playing r2 to do an action act. In other words: sendt+5 (agent A, r1, request(r1, r2, act)). When r1 receives a send from agent A asking for an act of r2, first it checks if the sender has the authority in the system to ask such an act, if so r2 acquires the goal to do act:
2(authoritysys (r , act) ⊃ [receive(r, agent A, request(r, r , act))]Gr (act)). Is important to underline that because role internals are public to other roles in the same system, it is always possible for r1 to check or modify r2’s goals. So, at time t+6 we have: receivet+6 (r1, agent A, request(r1, r2, act))
Merging Roles in Coordination and in Agent Deliberation
73
6. Now that r2 has updated its internal state (i.e. its goals) it must inform its player agent B: update statet+7 (r2, agent B) Where update state is modeled as in Rule 12 7. At time t+8 agent A decides to deact the role r1: sendt+8 (agent A, r1, deact) 8. Finally, at time t+9, r1 confirm the deact: confirm deactt+9 (r1, agent A)
7 Conclusions and Further Works In this article we merged two representative roles models in MAS by introducing a metamodel taken from Genovese [5] and adapting it to agents. In particular, we added representations of typical agents mental attitudes and a framework to deal with message passing. The model has been specialized in order to describe both public and private dimensions of roles (Boella, Dastani [[1,2]). Finally, we merged the two dimensions defining a group of actions together with their semantics and we modeled a running example to show a possible course of events. Further works point in two main directions: adapting the proposed metamodel to other roles approaches like Omicini [10], and introducing a formal proof theory of roles’ actions dynamics and related aspects starting from Baldoni et al [8].
References 1. Boella, G., Damiano, R., Hulstijn, J., van der Torre, L.: Acl semantics between social commitments and mental attitudes. In: Procs. of Workshop on Agent Communication (2006) 2. Dastani, M., van Riemsdijk, B., Hulstijn, J., Dignum, F., Meyer, J.J.: Enacting and deacting roles in agent programming. In: Odell, J.J., Giorgini, P., M¨uller, J.P. (eds.) AOSE 2004. LNCS, vol. 3382, pp. 189–204. Springer, Heidelberg (2005) 3. Zambonelli, F., Jennings, N., Wooldridge, M.: Developing multiagent systems: The Gaia methodology. IEEE Trans. of Software Engineering and Methodology 12(3), 317–370 (2003) 4. Colman, A., Han, J.: Roles, players and adaptable organizations. Applied Ontology (2007) 5. Genovese, V.: Towards a general framework for modelling roles. In: Normative Multi-agent Systems. Number 07122 in Dagstuhl Seminar Proceedings (2007) 6. Bellifemine, F., Poggi, A., Rimassa, G.: Developing multi-agent systems with a FIPAcompliant agent framework. Software – Practice And Experience 31(2), 103–128 (2001) 7. Baldoni, M., Boella, G., van der Torre, L.: Interaction between Objects in powerJava. Journal of Object Technology 6, 7–12 (2007) 8. Baldoni, M., Baroglio, C., Martelli, A., Patti, V.: Reasoning about interaction protocols for customizing web service selection and composition. Journal of Logic and Algebraic Programming, special issue on Web Services and Formal Methods 70(1), 53–73 (2007) 9. Boella, G., van der Torre, L.: The ontological properties of social roles in multi-agent systems: Denitional dependence, powers and roles playing roles. Artificial Intelligence and Law Journal (AILaw) (2007) 10. Omicini, A., Ricci, A., Viroli, M.: An algebraic approach for modelling organisation, roles and contexts in MAS. Applicable Algebra in Engineering. Communication and Computing 16, 151–178 (2005)
Planning Actions with Social Consequences Hsueh-Min Chang1 and Von-Wun Soo1, 2 1
Department of Computer Science, National Tsing Hua University 101, Section 2, Guangfu Road, Hsinchu 300, Taiwan 2 Department of Computer Science and Information Engineering, National University of Kaohsiung 700, Kaohsiung University Road, Kaohsiung 811, Taiwan {pchang,soo}@cs.nthu.edu.tw
Abstract. In an environment with multiple autonomous agents, the performance of an action may have effects on the beliefs and goals of the witnessing agents in addition to the direct effects. The awareness of such mental effects is critical for the success of a plan in multi-agent environments. This paper provides a formulation of social plans, and show that social planning can be done by including models of other agents' minds in the planning domain. A social planning agent is constructed based on automatic generation of PDDL (Planning Domain Description Language) domains from knowledge about other agents.
agents who ignore the mental effects of actions will be blind to many further social consequences, and often fail to produce a feasible plan. A closely related concept to social consequences is social actions, which Castelfranchi defines as those actions that “deal with another entity as an agent, i.e. an active, autonomous, goal-oriented entity.” [3] He further divides social actions into weak and strong ones. Weak social actions are based on beliefs, but not goals, about others' minds or actions. For instance, in case (i), if A believes that C intends to take the knife, A's taking another weapon is a weak social action, for it is based on A's own belief about C's goal but is not to change it. A strong social action, however, is directed by social goals, i.e. goals to change or maintain the mental states of others. For instance, in case (ii), if A chooses to take an alternative entrance or to drug the guard out of his consciousness, A is performing a strong social action directed by the social goal that D has no goal to block A's entrance. The model of social actions therefore distinguishes between the problem of plan coordination, which requires solely weak social actions, and the problem of social planning introduced in this paper. Social planning extends conventional planning by including social consequences as a part of its action effects, and thus requires strong social actions. In this work, we put the concepts into practice by presenting an implementation of social planning. Since social planning extends the state space to the internal states of agents, the planning agent must treat the internal mechanisms of other agents as part of the world causality. For example, an agent's perceiving a particular object causes its belief state to change, and the new belief state activates the motivational mechanism [11] to generate a new goal state, which can be translated into a new physical state of the world through actions. Essentially, the agent builds a model of other agents and simulates their mental processes in order to reason about what actions he can perform in order to bring about certain mental states of others. However, since mental states are not directly accessible, social planning will cause the well-known ramification problem ([2] and [8], for example), which states that actions can have indirect effects which is context-dependent and can be potentially unbounded. The implementation of the social planning agent takes a representation-oriented approach that leverages the expressive power of the Planning Domain Description Language (PDDL) [7][12] to describe agent perception, mental states, belief revision and desire generation in a limited but useful way. In particular, PDDL provides a derived predicates construct that can represent the deduction from physical states to mental states. The social planning agent translates its model of other agents to PDDL files, and uses one of the state-of-the-art general-purpose planners to do the actual planning. Details of domain translation are presented in Sect. 3. One of the potential applications of social planning is on the interactive drama domain, as is demonstrated in section 4, in which social planning agents are used to simulate a simplified version of Othello, a famous play by William Shakespeare.
2 Analysis of Social Planning Social planning can be viewed as basically classical single-agent planning enhanced with consideration on the social consequences of actions. The product of social planning is often (but not always) a social plan, which contains (a) actions by other agents and (b) actions by self that motivate these other agents to act. Although a social plan is a multi-agent plan, social planning is done from a single-agent perspective, utilizing actions by others only as a means to achieve its own goal. This sets social planning apart from many forms of multi-agent planning that assumes the agents are designed
76
H.-M. Chang and V.-W. Soo
to be a part of a cooperative system whose objective is to maximize the team utility, such as in [9]. In effect, social planning enhances classical planning in two ways. On one hand, active social planning aims to invoke others' goals that positively contributes to the plan, making it possible to construct plans for goals that an agent cannot achieve by itself. On the other hand, passive social planning aims to prevent a goal that is detrimental to the plan from being generated, and thus helps detect plans that work in single-agent settings but no longer works when other agents are present. Note that purely passive social planning can result in nonsocial plans because its ultimate aim is to trigger no beliefs and goals at all. This fact leads to subtleties about the identification of social actions, as will be shown in Sect. 2.2. In the remainder of the section, we will use a simple motivational, goal-oriented agent to illustrate the nature of social planning. 2.1 Formulation of Social Plans We begin by describing an agent with naïve rule-based belief revision and motivation functions. Although decidedly simple, the agent is motivated, i.e. can generate goals autonomously, and thus satisfies the basic requirement of social planning. Definition 1 (Description of an agent) Let P be the set of all atomic propositions. Let R be the set of first-order sentences. An agent ag1 is a 4-tuple (Bag1, Gag1, BRRag1, MRag1), where Bag1 P is ag1's beliefs, Gag1 P is ag1's goals, BRRag1 R is a set of belief revision rules, and MRag1 R is a set of motivation rules. The rules have the form x→p, where x is a first-order proposition and p P is the derived atom. A belief revision rule derives a belief from beliefs, while a motivation rule derives a goal from beliefs. The agent will gain perceptual input from the environment, update the belief set using BRRag1, and then update the goal set using MRag1. If any new goal is added, the agent searches for a plan for the goal. A social planning agent (denoted ag1 in this section) can thus have an internal model of another agent (denoted ag2) using the explicit representation. Definition 2 (Model of another agent) Agent ag1's model about another agent ag2 is a 4-tuple Mag2 = (Bag2, Gag2, BRRag2, MRag2), where Bag2 = {p | believes(ag2, p) Bag1} is ag1's model about ag2's beliefs, Gag2 = {p | hasGoal(ag2, p) Bag1} is ag1's model about ag2's goals. BRRag2 and MRag2 are the modeled belief revision functions and motivation functions, respectively.1 The social planning agent ag1 can then infer how a percept changes the internal state of ag2. However, ag1 also needs a perception model to estimate whether the effects of its action are actually perceived by ag2. Note that it is the environment, not the agents, who dictate the range of perception. We assume that for any event e P, the environment provides the truth value of the predicate perceives(ag, e). With the above models, now we can derive the mental changes as a result of an action. To calculate the mental effect, ag1 simulates an instance of ag2 by adopting ag2's beliefs, goals and mental state transition rules. Then ag1 gives the simulated ag2 some hypothetical 1
We adopt the naming convention that predicates are in lower case while functions are capitalized.
Planning Actions with Social Consequences
77
percepts which would result from the action, and observes the mental state change. It first applies the belief revision rules to derive the belief effects. Definition 3 (Belief effects of an action on one agent) Let P-Eff(action) be the set of physical effects of action. The set of belief addition by an action action on an agent ag2 is B-Add(action, Mag2) = {p | pP, p Bag2, {q | qPEff(action), perceives(ag2, q)} ∪ Bag2 ∪ BRRag2 ⊢ p}. The set of belief deletion is BDel(action, Mag2) = {p | p Bag2, {q | qP-Eff(action), perceives(ag2, q)} ∪ Bag2 ∪
BRRag2 ⊬ p}.
The belief effects thus include an add list and a delete list. The add list contains atoms that can be derived but was not originally believed, while the delete list contains ones that was originally believed but cannot be derived after an action is conducted. After the beliefs of ag2 are updated, motivation rules are applied to derive new goal effects. Definition 4 (Goal effects of an action on one agent) Let Bag2' be the belief set updated with B-Add(action, Mag2) and B-Del(action, Mag2). The set of goal additions by the action action on an agent ag2 is G-Add(action, Mag2) = {p | p P, p Gag2, Bag2' ∪ MRag2 ⊢ p}. The set of goal deletions isG-Del(action, Mag2) = {p | p Gag2, Bag2' ∪ MRag2 ⊬ p}. Now we turn to actions that consume rather than produce mental effects. For an agent who plans its own actions, the actions only need to satisfy the physical precondition. However, if the plan involves actions of other agents (which we will call foreign actions), the planning agent must motivate these agents to perform them. Riedl and Young [14] are among the first to propose this additional intentional requirement in planning, although their planner is not based on the perspective of an individual agent. In short, the intentional precondition of an action is that either the performer is self or the performer desires any of the physical effects of the action by other agents. Definition 5 (Intentional precondition) The intentional precondition of an action performed by agent ag is the disjunction of members of the set M-Precond(action, ag) = {goal(ag, p) | p P-Eff(action)} ∪ {self(ag)}. The total precondition is the thus conjunction of the physical precondition and the intentional precondition. Finally, we derive our definition of social plans as containing one or more foreign actions, whose intentional preconditions are supported by some other actions. For convenience, we adopt the convention in plan-space planning (e.g. POP) that there is a finish action whose precondition is the goal of the plan. Thus, if the goal of the plan is a social goal, finish is considered a foreign action, and the plan a social plan. Definition 6 (Social plan) Let Actions(plan) be the set of all actions in plan. social(plan) (a1, a2 Actions (plan)), ( (p, ag), (goal(ag, p) M-Precond(a1, Mag) p G-Add(a2, Mag))). 2.2 Relationship between Social Actions and Social Plans When analyzing a social plan, one may be tempted to regard it as the combination of three types of actions: social, non-social and foreign actions. While an action whose
78
H.-M. Chang and V.-W. Soo
effects include some intentional precondition of another action is clearly social, one that does not is not necessarily nonsocial. Here we provide two reasons why the such an action could still be social: 1.
2.
The action belongs to a subplan that causally supports a social action. Thus, the action still can be regarded as driven by a social goal, albeit indirectly. This rises the question of the degree of sociality. The action is chosen because alternatives are known to cause negative social consequences. The action could be regarded as driven by a social goal to keep another agent from generating a goal, and thus social. However, if the action is considered before the bad alternatives which are therefore never considered, can it still be said to be directed by a social goal?
In essence, Castelfranchi's social action is a concept based on the agent's subjective judgment about whether it is “dealing with another entity as an agent.” Whether an action is social depends on the agent's way of thinking, such as its choice of planning algorithm. Thus, it is simpler to regard social planning as incorporating social consequences than as producing social actions. On the other hand, a social plan is a structural concept. A plan is social because its structure involves other agents; how it is generated is irrelevant. 2.3 Social Planning as a Case of the Ramification Problem Social planning relies on mental effects, which are always indirect because agents have no direct access to others' minds [8]. They are also highly context-dependent; who the witnesses are and what they have in mind decide the mental effects of an action, as shown in Definition 3 and 4. These facts classify social planning as an instance of the long recognized ramification problem, which concerns the indirect consequences of an action. A variety of methods have been proposed to (partially) solve the ramification problem in planning, including deductive planning [2] and external theorem provers [8], among others. However, completely specifying the action ramifications can be difficult. In social planning, an action can cause mental effects, which motivate another agent to perform some action, which in turn can be perceived by yet another agent, and so on, resulting in a chain of reaction. An agent with knowledge about the mental states of a society can in principle predict the future of the society by modeling the agents' actions step by step while checking whether their interaction changes their goals and thus the course of their actions. The total effect of an action is thus the difference between how the society would evolve with and without that action. Luck and d'Inverno [10] called agents capable of modeling the society sociological agents. Unfortunately, the butterfly effects2 of an action could be broad in scope and infinite in time, making it difficult to find a sound and complete social planner. In the next section, we will introduce a method that compiles a limited representation of the social planning problem and then solve it with a sound and complete planner.
2
It is commonly said that a flap of wings of a butterfly could produce a tornado far away.
Planning Actions with Social Consequences
79
Fig. 1. The architecture of the social planning agent
3 Social Planning through Domain Compilation This section presents an implementation of social planning agents based on planning domain compilation. An assumption of automated planning is that the causality of the world is encoded in a domain-description language, which allows the specification of planning operators as triggers of state change. Therefore, planning can control the social consequences in multi-agent environments if the planning domain includes agency as part of world causality. Based on this observation, we have developed a social planning agent without resorting to a special-purpose planner3. Instead, the agent leverages the expressive power of PDDL to represent a social planning problem as a standard planning problem, and uses a general-purpose planner to solve it. Fig. 1 shows the architecture of the agent who plans socially with domain compilation techniques. Before planning, the agent compiles its beliefs and the ontology schema into PDDL files. A compatible planner then takes the files as input, plans, and returns the result to the agent. Domain compilation is based on the formulations of mental effects and preconditions in Sect. 2. It generates a partial representation of the social planning problem with an inherent bias towards active social planning. The remainder of this section will briefly introduce PDDL and describe the details of domain compilation into PDDL. 3.1 Introduction to PDDL McDermott et al. [12] created the first version Planning Domain Description Language (PDDL) in an attempt to provide a standard language for specifying planning problems in the International Planning Competition. PDDL has since been revised several times for the need of the competition. As a result, most mainstream planners now adopt PDDL documents as input format. Two PDDL documents are needed for a planning problem. The domain file defines the predicates, actions and their preconditions and effects for a planning domain. The
3
This is not to say that social planning cannot benefit from a special-purpose planner.
80
H.-M. Chang and V.-W. Soo
problem file contains the initial state defined using the predicates in the domain, and a goal statement. PDDL has a rich set of features, and most planners implement only a subset of PDDL. Barros and Musse [1] compiled a comparison of supported PDDL features for several modern planners, but did not include derived predicates, a features introduced in version 2.2 [7]. Among other features, derived predicates and conditional effects are the most crucial features for social planning because they enable flexible representation of context-dependent effects. Conditional effects allow the representation of effects which holds only if a certain condition holds before the action. On the other hand, derived predicates represent facts that can be derived from other predicates, but can never be direct effects of any action. A classic example is the above predicate in the blocks world, in which above(a, b) can only be derived from facts like on(a, c)on(c, b). Derived predicates are therefore a partial solution to the ramification problem. It is easy to see that elements of mental states also possess the properties of derived predicates because no action can directly infuse a belief in an agent; actions can only produce percepts, from which beliefs are derived. Optop [13] and Marvin [5] are two of the planners that provide the best support for derived predicates. 3.2 Automated Domain Compilation Fig. 2 lists the mapping between a motivated agent and the corresponding PDDL constructs that describe the agent. A domain compiler (DC) compiles the domain ontology and each part of the agent into PDDL. The generation of regular predicates and actions from the domain ontology is rather straightforward and thus is omitted. Specific processes required by social planning are discussed in the following. Representing beliefs and goals. Since PDDL is unable to represent nested literals like belief(Emilia, holds(Desdemona, Handkerchief)), DC must generate additional predicates like b_holds(Emilia, Desdemona, Handkerchief) for each predicate as a workaround. Similarly, DC generates p_holds for percepts and g_holds for goals. Perception. Any change to the mental state starts from perception. Thus, agents whose plan relies on foreign actions must have some control on the perception of other agents. We make the assumptions that agents have a perception range, and that they perceive all events happening within that range. Accordingly, an action produces percepts if and only if either of the following conditions holds: (a) the action is performed within the perception range of some other agent, (b) the action causes an object to enter the perception range of some agent. The performance of an action produces two types of events: an action event represents the performance of an action, while state events represent the changed state after the performance of the action. PDDL implementation of perception. Percepts are represented in PDDL as conditional effects of an action because an action successfully invokes percepts only when the geographic relationship between agents is valid.4 The following expression in terms of PDDL represents the conditional state event of performing drop(actor, thing, place): 4
Although one might be tempted to describe percepts as derived predicates dependent on geographic relations, doing so will cause existing percepts to disappear when the supporting geographic conditions no longer hold!
Planning Actions with Social Consequences
81
Fig. 1. The mapping between agents and the PDDL constructs (forall (?other) (when (and (in ?other ?place) (in ?actor ?place)) (and (p_in ?other ?thing ?place) (not (p_holds ?other ?thing ?actor)))) )
The action goto is more complex because it involves location change, and its corresponding code is too long to be listed here. Basically, performing a goto will cause the actor to perceive all (observable) predicates describing everything in the new place, and cause everyone in the new place to perceive all predicates describing the actor. Therefore, domain compilation has to list all predicates as the condition effects of goto. Belief revision and goal generation. Belief revision is the process of updating beliefs according to the percepts. There are two types of predicates for beliefs: action event belief and state belief, in correspondence with the two types of percepts. In addition to directly reflecting percepts, state beliefs can also be derived from existing ones using belief revision rules. For example, the social knowledge may describe agents as having a belief revision rule like the following one: loves(ag1, ag2) t, precious(t) gives(ag1, ag2, t) man(ag1) woman(ag2) Note that gives(ag1, ag2, t) is an action event belief, while others are state beliefs. However, a social planning agent adopts the rule not to revise its own beliefs, but to reason about how to activate it for other agents in order to change their minds. Therefore, the subjective rule needs to be transformed into a relative one, in which each literal becomes a belief of some particular agent. The resulting axiom is as follows: subject, belief(subject, loves(ag1, ag2) t, belief(subject, precious(t)) belief(subject, gives(ag1, ag2, t)) belief(subject, man(ag1)) belief(subject, woman(ag2)) DC uses the same procedure to compile motivation rules for describing goal generation. PDDL implementation of belief revision and goal generation. Beliefs and desires are described as PDDL derived predicates of the form (:derived P F), meaning that predicate P holds iff formula F holds. Here is an example of the above belief of love, specifying both sources of belief:
Below is a compiled motivation rule specifying that agents will generate goals to get precious things at the same place as his current location. (:derived (g_holds ?s ?t) (and (b_precious ?s ?t) (exists (?p) (and (in ?s ?p) (in ?t ?p))) )
Action selection. An agent will attempt to achieve the goals through actions. To represent the relationship between goals and actions, DC extends action preconditions according to the formulation of intentional precondition in Definition 5, which states that an action can be included in the plan only if some of its effects are desired by the performer. Although the use of intentional preconditions facilitates active social planning, it also leads to limitations in passive social planning. The reason is as follows. A satisfied precondition merely implies that the action is feasible, not that the action is necessary. This goes contrary to the fact that when an agent desires some effect of an action, the agent will perform it, unless there is an alternative action that provides the same effect. Thus, other agents’ actions whose intentional preconditions are accidentally satisfied during planning will be performed. A normal planner will simply choose not to include these actions in the plan if they are irrelevant or detrimental to the goal, but that they will still be performed is not changed. PDDL implementation of action selection. The PDDL is exactly modeled after Definition 5. For the drop action that has the physical effects in(thing, place) holds(actor, thing), the intentional precondition is (or (self ?actor) (g_in ?actor ?thing ?place) (g_not_holds ?actor ?thing ?actor))
Macro actions. Another problem is that the other agent may need more than one action to achieve the goal. To model this, social planning agent may have to plan for other agents and check interferences among all plans. Since plan coordination is outside the scope of this paper, we use macro actions to bypass the problem. Macro actions represent a sequence of actions to achieve a goal, like plan libraries. They have the same format as regular actions, except that they can only be performed by other agents. An example in our domain is steal-for, which represents walking to the victim, stealing something, walking to the master, and give it to the master. PDDL implementation of macro preconditions. The PDDL for macro actions is not different from that for regular actions in terms of perceptual effects. The only difference is that macro actions do not have the self disjunct in their intentional preconditions. Problem generation. Whenever a social planning agent chooses a goal to plan for, DC creates a PDDL problem file containing the goal and the agent's social and physical beliefs as the initial state. The translation process for beliefs to PDDL simply involves dumping all beliefs as PDDL literals. A physical belief has the form (holds Iago Handkerchief), while (g_holds Cassio Cassio Handkerchief) represents a social belief that Cassio desires to hold the handkerchief. With the domain and problem files, the planner can connect the actions to produce a plan.
Planning Actions with Social Consequences
83
4 Implementation Results The implementation is based on our ontology-based agent environment, which uses OWL models to describe the knowledge schema and action specification. Details about the system can be found in [4]. The domain compiler translates OWL models to PDDL files. Both the agents and the domain compiler are implemented in Java and use sockets to connect to McDermott's Optop planner, which runs on Lisp. 4.1 The Othello Scenario We draw our inspiration from Shakespeare's Othello, a prime example of shrewd social planning. A vastly simplified version of it is used as the scenario for demonstration. Here is our mundane version: A soldier Iago have long resented General Othello and plotted to make him suffer by having him kill his own wife, Desdemona. To carry out his plan, Iago first asked his wife Emilia in the garden to steal a handkerchief Othello gave Desdemona. Emilia stole it from Desdemona and gave it to Iago. Iago then deliberately planted the handkerchief in the lieutenant Cassio's room, expecting that Cassio would pick it up, and Cassio did it. Othello, under the advice of Iago, then paid a visit to Cassio. Seeing the handkerchief on Cassio, he was enraged by jealousy. He killed Cassio immediately and ended up going to Desdemona's room to kill her as well. All the characters are modeled as autonomous motivated agents, but Iago is the only social planning agent among them. The scenario contains 9 objects, 6 actions, 1 macro action, and 16 basic predicates, each of which has 3 extended versions: belief, goal and precept. The actions are goto, pickup, drop, kill, request_in, and request_holds, the latter two being communicative actions representing a request to adopt a goal to be in somewhere or to hold something. The macro action is steal_for. One belief predicate, b_loves, is derived from a motivation rule of folk psychology basically stating that when an agent C finds a gift given to person A to be on person B of an opposite sex, C believes that A loves B. Two goal predicates, g_holds and g_dead, are derived from two motivation rules, greed and jealousy. The greed rule states that an agent will desire to take away precious things in his residing location, and the jealousy rule states that an agent will desire to kill his wife and her lover, if the lover is not the agent himself. 4.2 Results and Performance In the simulations, both active and passive social planning plays a part in Iago's plan. In a simulation run in which Iago's goal was hasGoal(Othello, dead(Desdemona)), Iago did produce a plan, but this plan included an action to carry the handkerchief to Othello's home, resulting in the deaths of both Desdemona and Iago himself, the latter being an undesirable side effect. To correct this, the goal is conjoined with hasGoal(Othello, dead(Iago)) to enforce the passive social planning to remove the side effect. On a 3.0 GHz Pentium 4 computer running CLISP, Optop takes 10.2 seconds to find a correct plan. The domain compiler needs to be ported to work with Marvin [5], whose support on conditional effects is limited. We have not finished the porting, but our initial experience is that Marvin takes excessive time in preprocessing the domain. The output of Optop is presented in the following with adjusted indentation. However, it comes up with a plan that is slightly different from the original scenario in that Iago requests Cassio to visit Othello, not vice versa.
84
H.-M. Chang and V.-W. Soo
(#) ;
5 Conclusion When multiple motivated agents inhabit the same environment where they can see each other, effects of an action will enter the witnesses' minds and change their beliefs and goals and could thus initiate a series of reactions. Since actions may have social ramifications, an agent can no longer be oblivious of others while making plans. Instead, they should use actions to motivate others to conduct parts of the plan, and avoid actions with negative social consequences. The formulation of social plans takes into consideration of the dependence relation between actions of self and actions of others in a plan. Although a sound and complete social planner is very difficult to produce, we show that a reduced social planning problem can be represented using a standard domain description language. The implementation of social planning agents demonstrates that social planning through planning domain compilation can be done with a general purpose planner. Since social planning makes no assumption about whether agents are cooperative or competitive, improved understanding of the problem, its logical characterization and its computational properties could shed lights on how human-like agents interact socially in general.
Acknowledgment This work is supported by the National Science Council of Taiwan under grant number NSC 96-2628-E-007-044-MY3.
References 1. Barros, L.M., Musse, S.R.: Planning algorithms for interactive storytelling. Comput. Entertain. 5 (2007) 2. Bibel, W.: Let’s Plan it Deductively! Artif. Intell. 103, 183–208 (1998) 3. Castelfranchi, C.: Modeling Social Actions for AI agents. Artif. Intel. 103, 157–182 (1998) 4. Chang, P.H.-M., Chien, Y.-H., Kao, E.C.-C., Soo, V.-W.: A Knowledge-Based Scenario Framework to Support Intelligent Planning Characters. In: Panayiotopoulos, T., Gratch, J., Aylett, R.S., Ballin, D., Olivier, P., Rist, T. (eds.) IVA 2005. LNCS, vol. 3661, pp. 134– 145. Springer, Heidelberg (2005) 5. Coles, A.I., Smith, A.J.: Marvin: Macro-actions from reduced versions of the instance. In: IPC4 Booklet, Fourteenth International Conference on Automated Planning and Scheduling (2004) 6. Cox, J., Durfee, E.H.: An Efficient Algorithm for Multiagent Plan Coordination. In: Proceedings of the 4th International Joint Conference on Autonomous Agents and MultiAgent Systems (AAMAS 2005), pp. 828–835. ACM Press, NY (2005)
Planning Actions with Social Consequences
85
7. Edelkamp, S., Hoffmann, J.: PDDL2.2: The Language for the Classical Part of the 4th International Planning Competition. Technical Report 195. Albert-Ludwigs-Universität, Institut für Informatik, Freiburg, Germany (2004) 8. Field, D., Ramsay, A.: How to Change a Person’s Mind: Understanding the Difference between the Effects and Consequences of Speech Acts. In: Proceedings of 5th Workshop on Inference in Computational Semantics (IcoS-5), Buxton, England, pp. 27–36 (2006) 9. Grosz, B., Hunsberger, L., Kraus, S.: Planning and Acting Together. AI Magazine 20, 23– 34 (1999) 10. Luck, M., d’Inverno, M.: Plan Analysis for Autonomous Sociological Agents. In: Castelfranchi, C., Lespérance, Y. (eds.) ATAL 2000. LNCS, vol. 1986, pp. 182–197. Springer, Heidelberg (2001) 11. Luck, M., d’Inverno, M.: Motivated behaviour for goal adoption. In: Zhang, C., Lukose, D. (eds.) DAI 1998. LNCS, vol. 1544, pp. 58–73. Springer, Heidelberg (1998) 12. McDermott, D., Ghallab, M., Howe, A., Knoblock, C., Ram, A., Veloso, M., Weld, D., Wilkins, D.: PDDL - The Planning Domain Definition Language. Technical Report CVC TR-98-003/DCS TR-1165. Yale Center for Computational Vision and Control (1998), http://www.cs.yale.edu/~dvm 13. McDermott, D.: The Optop Planner. In: IPC4 Booklet, Fourteenth International Conference on Automated Planning and Scheduling (2004) 14. Riedl, M.O., Young, R.M.: An Intent-Driven Planner for Multi-Agent Story Generation. In: Proceedings of the Third International Joint Conference on Autonomous Agents and Multi-Agent Systems (AAMAS 2004), pp. 186–193. IEEE Computer Society, Washington (2004) 15. Rowe, N., Andrade, S.: Counterplanning for Multi-agent Plans using Stochastic MeansEnds Analysis. In: Proceedings of the IASTED Artificial Intelligence and Applications Conference, pp. 405–441 (2002)
Layered Cooperation of Macro Agents and Micro Agents in Cooperative Active Contour Model Noriko Matsumoto, Norihiko Yoshida, and Shuji Narazaki Division of Mathematics, Electronics and Informatics Saitama University, Saitama 338-8570, Japan [email protected]
Abstract. We have proposed Multi-Snake, which realizes boundary detection in image recognition with the layered cooperation of micro agents and macro agents. Cooperation in a set of micro agents constructs the behavior of a macro agent, and cooperation of the micro agents are integrated to cooperation of the macro agents. This mechanism makes the application more dynamic and flexible. Our previous proposals dealt with cooperation between some macro agents of the same kind. This paper focuses on the cooperation of macro agents of different kinds: sensor-based macro agents and model-based macro agents. We show that our proposal makes estimation improved and more robust. We verify the effectiveness of our proposal through some experiments using artificial images and real images.
1
Introduction
Image recognition and feature extraction are still difficult to solve although there have been many proposals applying various techniques including agents techniques. Accurate boundary detection in an image is one of the most important problems in image recognition, however it is also still difficult. One of the most popular boundary detectors is Active Contour Model “Snake”, proposed by Kass et al. [1]. The principle of Snake is an energy-minimizing spline for estimating the closest contour of a target object in an image gradually from an initial contour. This principle can be considered as a boundary detector realized by cooperation in an aggregate of many micro agents, that is, the contour is a single macro agent which consists of a set of micro agents. There have been some proposals concerned with cooperation of macro agents. In this paper, we discuss layered cooperation which integrates the cooperation of the micro agents, who construct a macro agent, to the cooperation of the macro agents. In this way, we can produce the macro cooperation from the micro cooperation. The layered cooperation make this application more dynamic and flexible. We have proposed Multi-Snake, the improvement of Snake in this manner. This proposal so far has been concerned with cooperation of macro agents of the same kind, and we verified its effectiveness [2,3,4]. Our another paper [5] presented some preliminary idea on cooperation of macro agents of different kinds, namely, sensor-based macro agents and model-based macro agents. In A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 86–97, 2009. c Springer-Verlag Berlin Heidelberg 2009
Layered Cooperation of Macro Agents and Micro Agents
(a) Operation at each control point
87
(b) Macro agent composed of micro agents
Fig. 1. Principle of Snake
this paper, we investigate some fundamental aspects on cooperation of macro agents of different kinds based on cooperation of micro agents.
2
Active Contour Model, Snake
Snake is a deformable contour v(s) = [x(s), y(s)], s ∈ [0, 1] that moves to minimize an energy function Esnake . The energy Esnake consists of an internal force Eint (v(s)) and an external force Eext (v(s)). Eint (v(s)) indicates elasticity and smoothness of the contour, and Eext (v(s)) is derived from the image along the contour v(s) so as to be attracted to a certain edge. Snake is formulated as minimizing a spline of the energy function as follows: Esnake = {Eint (v(s)) + Eext (v(s))}ds (1) Eint (v(s)) = (α|vs (s)|2 + β|vss (s)|2 ) / 2
(2)
Eext (v(s)) = −γ|∇I(v(s))|2
(3)
where vs (s) and vss (s) denote the first and second derivative of the contour with respect to s, ∇I is image intensity gradient, and each α, β, and γ are weight coefficients. It means the value of this energy function gets smaller, if the shape of the closed contour is more circular, if the circumference length is shorter, or if the intensity gradient is larger. However, this tendency depends on the parameter settings. In general, Snake transforms the contour to minimize Esnake from an initial closed contour provided by a user, and searches the target boundary. Snake calculates the changes of the function value at each point, namely each control point, on the closed contour, and moves the control point to a direction whose energy gradient is the steepest as shown in Figure. 1(a). Therefore, it can be considered that each control point is a micro agent and, the contour is a single macro agent as an aggregate of the micro agents (Figure. 1(b)). Snake searches the target boundary sequentially from the initial contour, so that it tends to be influenced extremely by the initial contour. Snake also has
88
N. Matsumoto, N. Yoshida, and S. Narazaki
a drawback of strong dependence to the local information, which is intensity gradient along the closed contour, as expressed in the formula (3). Hence, the detection accuracy becomes worse when the target image consists of complicated features. In order to address this kind of issue, various methods have been proposed using some macro agents distributed in space, such as a method dividing an image into several uniform regions beforehand [6], a method applying two Snakes simultaneously to the target region and the background region [7], and a method making some Snakes compete or cooperate [8,9]. However, they require strict positioning of many sample points for the initial contour(s).
3
Multi-Snake
The original Snake has some drawbacks of its strong dependence on the parameters, the target features and the target image as mentioned above. In order to address these drawbacks, we have proposed Multi-Snake, an improvement of Snake by applying the agent technique, with the layered cooperation mechanism [2,3,4]. This method uses several macro agents in parallel, each of which runs its own Snake, to detect the boundary of a single target object. Each macro agent does a different estimation being based on different solution space, using different parameters, or being applied to closely-related different images. This cooperation of macro agents are realized by the cooperation of micro agents, that is, Multi-Snake has two layers as shown in Figure 2 (a) and (b). Some micro agents, which are located at the same control point on the contour, exchange their intermediate contour estimation with each other periodically, and adjust their own contour estimations. This approach realizes robust estimation against noises and textures, and improves estimation qualities.
(a) In the macro layer
(b) In the micro layer
(c) Decision of new control points Fig. 2. Cooperation mechanism of Multi-Snake
Layered Cooperation of Macro Agents and Micro Agents
89
When the gradient (the first derivative) of the energy function calculated in a certain macro agent gets small, this means the agent gets near to a local minimum or the global minimum. However, if the gradient in another agent still keeps large at the same control point, it is supposed that the agent is not at the global minimum but at a local minimum. Accordingly, in the micro agent layer, each corresponding micro agent exchanges with each other the direction which gives the largest energy gradient among its eight-neighbors, and follows the direction which gives the largest gradient among the values from all the micro agents. This scheme is shown in Figure 2(c). This exchange takes place at every moment when the micro agents move control points. This cooperation of the micro agents layer realizes the cooperation of the macro agents layer.
4
Model-Based Snake
As mentioned above, Multi-Snake has two layers as shown in Figure 2 (a) and (b), and we have verified its effectiveness, but using macro agents of the same kind so far. To improve the estimation quality even more, we discuss cooperation of macro agents of different kinds [5], namely sensor-based macro agents and model-based macro agents. First, we introduce model-based Snake. Snake is prone to complex shapes, concave boundaries, and pointed boundaries. In particular, the concave boundaries cannot be detected by adjusting its parameters only, and this problem is still left as a big problem. This is caused by the energy potential as shown in Figure 3 (a), in which arrows express the forces to attract Snake to the target. A user must place an initial contour on the region very close to the target concave region so as to make Snake converge to the concave. Once Snake escapes from the energy potential to attract to the concave, it is inevitable to transform into a straight line as shown in Figure 3 (b). This is the core of the problem. In order to solve this kind of problem, we have an observation that the initial contour placed by a user implies (or can be given easily) some information of the target shape. There have been some methods using a model as information of the target shape; for example, a proposal determines Eint which is suitable for the target shape by presenting the prototype beforehand [10,11], or another proposal determines a range of acceptable shape divergence by estimating the average shape of target and the unevenness of shape [12,13,14,15,16]. In particular, Matsuzawa [16] adopts the symbolized information, such as “corner”, “curve”, “segment”, and “arc”, which is obtained from the shape of initial contour. Based on this symbolized information, this method assigns these shape information to each control point of Snake respectively, and makes control of Snake reflect the shape knowledge of target region. This approach has an advantage of its ability which can include the model with the process of boundary detection without giving up Snake’s advantages of the ability of general-purposeness and handiness. Based on this approach, we have considered modeling the fragment shapes
90
N. Matsumoto, N. Yoshida, and S. Narazaki
(a) Potential
(b) Snake convergence
(c) Our idea
Fig. 3. Around a concave boundary region
of the initial contour to bring in the model to the control of Snake. Matsuzawa’s symbolization cannot describe the fine features, although control Snake by the target model acquired from the shape information in an initial contour. Hence, we use the curvature at each region more concretely. Our model-based Snake, which we call Curvature Snake, selects the feature points from the initial contour, and models the curvature among the feature points. We bring in this curvature to Snake as a model. Curvature Snake uses the energy function as follows: (4) ECS = {Eint (v(s)) + Eext (v(s)) + Emodel (v(s))}ds 0 − Ecurv | Emodel =δ|Ecurv
(5)
0 is the initial curvature calculated on a feature point on the initial where Ecurv contour. These initial feature points are placed on the contour at an equal interval (Δs) as shown in Figure 4 (a). Ecurv is the current curvature calculated on a feature point on the current contour at each moment. When the current curvature approximates to the model (the initial curvature), Emodel gets smaller, therefore Curvature Snake can hold the information from the initial contour. When we notice the details of transformation of Snake, namely the movement of the control points, there is a case that the control points move along the contour. Therefore, this method needs a mechanism corresponding between a model and the current feature point. At each moment, the current feature point is selected from the control point which is on the base-line. Curvature Snake calculates ECS (Form.(4)) only when it is on the feature points. When on the other control points, it calculates the original energy function only. There would be various kinds of base-lines, such as the normals in the initial contour, or the line drawn from a gravity center of the initial contour; here we use the latter, because the feature points gather to concave region according to the convergence as shown in Figure 4 (b). Now, we describe the calculation of curvature from the feature points. The initial feature points is determined as the points dividing the initial contour at an equal interval Δs (Figure 4(a)). From these feature points, the calculation uses the first and second derivative as follows:
Layered Cooperation of Macro Agents and Micro Agents
91
(a) Initial feature points and base-lines (b) Current feature points Fig. 4. Principle of Curvature Snake
dxi xi+1 − xi−1 d2 xi xi+1 − 2xi + xi−1 ≈ ≈ ds 2Δs ds2 Δs2 2 dyi yi+2 − yi−1 d y y − 2yi + yi−1 i i+1 ≈ ≈ ds 2Δs ds2 Δs2 dxi d2 xi where and is called xs , xss , the curvature ki of feature point Si is deds ds2 fined as follows: yss xs − xss ys ki= 3 (x2s + ys2 ) 2 On the contrary, the current feature point in the detection process at time t cannot use the same equation as mentioned above, because its interval between the feature points is different respectively (Fig. 4(b)). However, the current curvature need not be completely exact, so that it is approximated during the detection process as follows: dxi xi+1 − xi−1 d2 xi xi+1 − 2xi + xi−1 ≈ ≈ ds Δsi−1 + Δsi ds2 Δsi−1 · Δsi 2 dyi yi+2 − yi−1 d yi yi+1 − 2yi + yi−1 ≈ ≈ ds Δsi−1 + Δsi ds2 Δsi−1 · Δsi where Δsi is the arc between the feature point Si and Si+1 as shown in Fig. 4(b).
5
Curvature Multi-Snake
Now we propose the layered cooperation of sensor-based agents and model-based agents, namely, the original Snake agents and Curvature Snake agents. We call this cooperation “Curvature Multi-Snake (CMS)”. Multi-Snake has two layers as shown in Figure 2 (a) and (b), and CMS also has two layers. In CMS, the original Snake cooperates with Curvature Snake at the feature points, because there are cases that the only Curvature Snake cannot keep the smoothness of the contour depending heavily on the shape of initial contour at each feature point.
92
N. Matsumoto, N. Yoshida, and S. Narazaki
At a feature point which crosses the base-line, each micro agents exchanges with each other the direction which gives the largest energy gradient among its eight-neighbors, and follows the direction which gives the largest gradient among the values from all the micro agents in the same manner of Multi-Snake. The differences between Multi-Snake and CMS is only the occasion of when the cooperation is happen. CMS makes agent cooperate only at the feature points. The overall detection procedure of CMS is as follows: Set an initial contour (by user) Find the center of initial contour Divide the initial contour at an equal interval (Δs) Repeat for each control point on the contour If the control point is on the base-line Make Original Snake cooperate with Curvature Snake Else Find a new control point using the original energy Until converged. Preserving the model at each feature point by Curvature Snake, CMS is able to overcome the difficulty of boundary detection without giving up Snake’s advantage of the smooth result contour by this cooperation performed only on the feature points.
6
Experiments
In order to verify the effectiveness of our proposal, we performed several experiments about Curvature Snake and CMS. In these experiments, we used artificial images (100 × 100 pixels) and real images (640 × 480 pixels) which have 256 levels of intensity for each color at each pixel. First, we show results of experiments using some artificial images. Figure 5 (a) presents the initial contour in a solid black curve, (b) is the results of the original Snake with (α, β, γ) shown under the pictures, and (c) is the results of the single Curvature Snake with (α, β, γ, δ). In this experiment, Curvature Snake does not cooperate with the original Snake on the feature points, but uses the energy function ECM only. On other control points, it performs the same as the original Snake. The purpose of this is to verify the effectiveness of the model force Emodel added to the energy function. In this picture, the result of boundary detection are shown in a solid black curve, the red pixels on each contour are the feature points. The number of feature points is thirty. As shown in (c), Curvature Snake failed to exhibit enough accuracies even with preserving shape. This is caused by the lack of desired effects of the model, namely the lack of appropriate correspondence between initial feature points and the current feature points. This result means that the single Curvature Snake has the target shape which cannot coordinate with the feature points as desired by a user. It depends on the target shape or the movement of control points. Then, we show how CMS influences the result accuracy when it makes original Snake of parameters (α, β, γ) cooperate with Curvature Snake of parameters
Layered Cooperation of Macro Agents and Micro Agents
(0.0, 0.0, 0.0, δ). In this case, Curvature Snake is not affected by Eint and Eext , so that the micro agents of original Snake cooperate with a model-based micro agent which depends on Emodel only at each feature point. (d) is the result of this case, and the parameters are shown under the pictures. This result shows the improvement of the accuracy of detection boundary. Additionally, we performed the experiments in which original Snake is cooperated with Curvature Snake with parameters (α, β, γ, δ) instead of (0.0, 0.0, 0.0, δ). (e) exhibits enough accuracies as well as the experiments as in (d). In these experiments of CMS, we checked which agents were chosen, and confirmed tendency to choose the original Snake agent as its convergence progressed. In the experiments of the single Curvature Snake (Figure 5 (c)), the process of boundary detection is performed with parameters (α, β, γ, δ), so that Snake is certainly influenced by the model force. On the contrary, the process in CMS ((d),(e)) chooses better suited agents according to both the contour shape and the image force at each moment. This means that CMS adjusts the influence of the model-based agents dynamically, consequently reduces the undesired transformation of the contour in concave regions, and at the same time can detect more smooth boundaries than the single Curvature Snake. However, CMS cannot exhibit enough accuracy with all parameter sets. There are some cases which failed to detect the concave region, for example we show that results in Figure 6 (a) and (b). CMS can improve its accuracies by increasing macro agents of original Snake whose parameter sets are different respectively. We performed the experiments of CMS employing three agents, namely two sensor-based agents and one model-based agent. Figure 6 (c) show the result, and through this experiment, we confirmed that CMS using three agents improved the stability of concave detection and also the dependence on a parameter set by reducing a tendency of boundary detection that parameters have. We also performed another experiment employing three agents, that is one sensor-based agent and two model-based agents, and (d) shows its result. This experiment also improved the accuracy of boundary detection.
Fig. 6. The results of CMS with 3 agents Table 1. Performance Point num Original Snake CMS 10 CMS 20 CMS 30
Loop 103 98 78 62
Time Time/Loop 0.157 0.00152 0.422 0.00431 0.359 0.00460 0.328 0.00529
We adopted thirty feature points in the above-mentioned experiments. Here we show the effect of changing the number of feature points for the detection result. Figure 7 is the result of CMS using the different number of feature point, and Table 1 shows the performance of this experiment. In this regard, we use two agents, one is an original Snake agent with parameter (0.6, 0.4, 1.0) and another is a Curvature Snake agent with parameter (0.6, 0.4, 1.0, 1.0). As in Figure 7, the number of feature points affects its result of boundary detection. In Table 1, “Loop” denotes the number of loops and “Time/Loop” denotes the execution time of one loop. The Time/Loop of CMS is about three times longer than that of original Snake. It is because of both of curvature calculation on each feature point and calculation of Emodel and ki . Next, Figure 8 is the result of CMS using real images. (a) presents the intensity of the image by color gradation, and (b) is the initial contour in black. In this case, the number of control points is 1784 and the number of feature points is 180. (c) is the result of original Snake of parameter (0.6, 0.4, 0.5), and (d) is the result of CMS that original Snake of parameter (0.6, 0.4, 0.5) cooperating with Curvature Snake with (0.6, 0.4, 0.5, 1.5). Even with CMS, it failed to exhibit enough accuracies with these parameter sets. This is caused by the undesired correspondence between the initial and current feature points. The target boundary has some concave regions, and this kind of targets needs some additional technique to correspond between them. Another reason is that the energy potential in the concave region is small. The intensity gradient is small
Layered Cooperation of Macro Agents and Micro Agents
95
(a) Using 10 points (b) Using 20 points Fig. 7. Effects of the Number of Representative Points
(a) Intensity of target image
(b) Initial contour
(c) S: (0.6, 0.4, 0.5)
(d) S: (0.6, 0.4, 0.5) + CS: (0.6, 0.4, 0.5, 1.5)
(e) S: (0.6, 0.4, 1.0)
(f) S: (0.6, 0.4, 1.0) + CS: (0.6, 0.4, 1.0, 1.5)
Fig. 8. CMS: (α, β, γ)+(α, β, γ, δ) applied to real image
between the target region and the background, and as a result, the force to attract to concave becomes smaller. Because the energy potential is formed by the intensity gradient and the parameters. When γ is set larger, (e) is the result of original Snake with (0.6, 0.4, 1.0), and (f) is the result of CMS that it cooperates with Curvature Snake with (0.6, 0.4, 1.0, 1.5). Through this experiment, CMS realized the improvement of stability in the concave region, and we confirmed the improvement of the accuracy.
7
Conclusions
Snake’s principle can be considered as a boundary detector realized by cooperation in an aggregate of many micro agents, and according to this point, we have
96
N. Matsumoto, N. Yoshida, and S. Narazaki
proposed Multi-Snake, the layered cooperation which integrates the cooperation of the micro agents, who construct a macro agent, to the cooperation of the macro agents. However, the previous Multi-Snake used some macro agents of the same kind. In order to improve its estimation quality, and to solve Snake’s vulnerability with concave regions, this paper proposed Curvature Snake, which adopts a curvature-based shape model based on information out of the initial contour, and also proposed Curvature Multi-Snake, layered cooperation of micro agents and macro agents of different kinds. We verified its effectiveness through some experiments. In the experiments of the single Curvature Snake, the drawbacks of Snake to the concave region was improved. However, there is a shape which could not exhibit enough accuracy even with Curvature Snake. This was caused by the undesired correspondence between the feature points. On the other hand, in the experiments of CMS improved the accuracy. In the detection process of CMS, we confirmed a tendency that CMS chose the agent of the original Snake in the last stages of convergence. This tendency shows that this layered cooperation of macro agents of different kinds is not only improving its accuracy but also this method is more dynamic and flexible. CMS can detect the boundary without giving up Snake’s advantage which can obtain the smooth boundary. In general, such kind of the complex shapes used in the experiment of real image is difficult for Snake, and imposes a great deal of labors on a user to set the initial contour and parameters suited for the target shape. Curvature Snake and CMS also need to set the initial contour with the contrivance for preserving the target shape. That is, the initial contour needs to be set to preserve the target shape such as concavity and convexity. On the other hand, the parameters is set through a certain amount of trial and error. Curvature Snake has the force to preserve the shape, so the internal energy Eint lessen its grip on Snake. As a result, Curvature Snake and CMS have an advantage that makes easier to determine parameters for users. If the initial contour includes concave regions, there is a case of one-to-many correspondence, namely an initial feature point and some current feature points. We dealt with this issue by selecting the feature point whose interval is shortest between the initial feature point and the current feature point. If Curvature Snake and CMS failed to detect the boundary because of this kind of correspondence, it will be preventable issue by cooperating with Curvature Snake of different number of feature points.
References 1. Kass, M., Witkin, A., Terzopoulos, D.: Snakes: Active contour model. Int.J. Computer Vision 1(4), 321–331 (1988) 2. Matsumoto, N., Yoshida, N., Narazaki, S.: Active Contour Model with Decentralized Cooperative Processing. IIEEJ 34(6), 747–752 (2005) 3. Matsumoto, N., Yoshida, N., Narazaki, S.: Cooperative Active Contour Model and Its Application to Remote Sensing. In: Proc. ACM 21st Annual Symp. on Applied Computing, pp. 44–45 (2006)
Layered Cooperation of Macro Agents and Micro Agents
97
4. Matsumoto, N., Yoshida, N., Narazaki, S.: Improvement of Active Contour Model with Decentralized Cooperative Processing and Its Application to Remote Sensing. International Journal of Knowledge-Based and Intelligent Engineering Systems 11(3), 169–179 (2007) 5. Matsumoto, N., Yoshida, N., Narazaki, S.: Curvature Multi-Snake: Cooperative Snakes with Curvature-Based Simple Modeling. In: Proceedings of International Workshop on Advanced Image Technology 2007, pp. 632–637 (2007) 6. Etoh, M., Shirai, Y., Asada, M.: Active Contour Extraction Based on Region Descriptions Obtained from Clustering. IEICE Transactions(D-II) J75-D-II(7), 1111– 1119 (1992) 7. Zhu, S., Yuille, A.: Region competition: Unifying snakes, region growing, and bayes/MDL for multiband image segmentation. IEEE Trans., PAMI 18(9), 884–900 (1996) 8. Wada, T., Nomura, Y., Matsuyama, T.: Cooperative Distributed Image Segmentation. IPSJ Journal 36(4), 879–891 (1995) (in Japanese) 9. Matsuzawa, Y., Abe, T.: Region Extraction Using Competition of Multiple Active Contour Models. IEICE Transactions(D-II) J83-D-II(4), 1100–1109 (2000) (in Japanese) 10. Staib, L., Duncan, J.: Boundary finding with parametrically deformable models. IEEE Trans. PAMI 14(11), 1061–1075 (1992) 11. Umeyama, S.: Contour extraction using a complex autoregressive model. Systems and Computers in Japan 28(1), 66–73 (1997) 12. Cootes, T., Taylor, C., Cooper, D., Graham, J.: Active Shape models – Their Training and Application. Computer Vision and Image Understanding 61(1), 38– 59 (1995) 13. Lai, K., Chin, R.: Deformable contours: Modeling and extraction. IEEE Trans. PAMI 17(11), 1084–1090 (1995) 14. Wang, Y., Staib, L.: Boundary finding with prior shape and smoothness models. IEEE Trans. PAMI 22(7), 738–743 (2000) 15. Olstad, B., Torp, A.: Encoding of a priori information in active contour models. IEEE Trans. PAMI 18(9), 863–872 (1996) 16. Matsuzawa, Y., Abe, T., Kumazawa, I.: Active Contour Model Using Shape Information Obtaned from Initial Contour. IEICE Transactions(D-II) J85-D-II(9), 1436–1445 (2002)
Contextual Agent Deliberation in Defeasible Logic Mehdi Dastani1 , Guido Governatori2, Antonino Rotolo3 , Insu Song2, and Leendert van der Torre4 2
1 Department of Information and Computing Sciences, University of Utrecht School of Information Technology and Electrical Engineering, The University of Queensland 3 CIRSFID, University of Bologna 4 Universit´e du Luxembourg, Computer Science and Communication
Abstract. This article extends Defeasible Logic to deal with the contextual deliberation process of cognitive agents. First, we introduce meta-rules to reason with rules. Meta-rules are rules that have as a consequent rules for motivational components, such as obligations, intentions and desires. In other words, they include nested rules. Second, we introduce explicit preferences among rules. They deal with complex structures where nested rules can be involved.
1 Introduction Logic is used in agent oriented software engineering not only for specification and verification, but also for programming deliberation and meta-deliberation tasks. For this reason, Defeasible Logic (DL) [1,2] has been extended with, amongst others, rule types, preferences [7], and actions [5,7]. In rule based cognitive agents, for example in defeasible logic, detailed interactions among cognitive attitudes like beliefs, desires, intentions and obligations are represented by rules, like the obligation to travel to Paris next week leading to a desire to travel by train (r1 ), or by preferences, such that if the desire to travel by train cannot be met, than there is a desire to travel by plane (p1 ). Patterns of such interactions are represented by rule priorities (obligations override desires or intentions – for social agents) [3,9] rule conversions (obligations behave as desires – for norm internalizing agents) [7], and so on. As interaction among mental attitudes becomes more complicated, the new challenge in agent deliberation languages is the coordination of such interactions. For example, at one moment an obligation to travel may lead to the desire to travel by train (r1 ), whereas at another moment it may lead to a desire to travel by plane (r2 ). Such coordination may be expressed by making the context explicit in rules r1 and r2 , and defining when rule r1 has higher priority than rule r2 , or it can be defined as a combination of rule r1 and preference p1 [5]. In this paper we raise the question how, as a further sophistication to coordinate the interaction among mental attitudes, to define the proof theory of nested rules (see [11] for a general theory of nested rules, [6] for theory of nested rules with time) and preferences among rules. Surprisingly, this complex language gives us just the right expressive power to describe a wide class of interaction phenomena: the rule leading to desire to travel by train may be preferred to the rule leading to a desire to travel by plane (r1 is preferred to r2 ), maybe as a second alternative, or the train rule may even A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 98–109, 2009. c Springer-Verlag Berlin Heidelberg 2009
Contextual Agent Deliberation in Defeasible Logic
99
be replaced by the plane rule (r1 into r2 ), maybe due to experienced train delays. The new language can be used to describe a new class of patterns of the coordination of interaction, for example when social agents turn into selfish agents, maybe when the agent does not have sufficient resources. Due to space limitations, we focus only on the formal aspects of the new logic. Finally, the definitions of the deliberation logics developed here are much more complex than the definitions of temporal logics traditionally used in agent based software engineering for specification and verification, since they contain rules, preferences, nonmonotonic proof system, and so on. However, whereas these temporal logics have a relatively high computational complexity, deliberation logics have to be efficient – with at most linear complexity (in the number of rules). Moreover, interaction patterns in such temporal logics have focussed on a relatively small class of agent types such as, for example, realisms and commitment strategies in BDI-CTL [4,10], whereas a much broader class has been studied in the more expressive deliberation logics. This paper is organized as follows. We first give some general intuitions of how the logical system works and define the formal language we use to contextualise the deliberation of cognitive agents. A running example will help us illustrate the system. The second step consists in describing the logical machinery for reasoning about contextualised agents. This requires to provide proof procedures to derive goals (desires, intentions, and obligations) as well as to derive rules for proving goals. A further example concludes the paper.
2 Contextualising Cognitive Agents The basic deliberative process uses rules to derive goals (desires, intentions, obligations) based on existing beliefs, desires, intentions and obligations (beliefs concern the knowledge an agent has about the world: they are not in themselves motivations for action). Contextualising the deliberation requires to provide the agent with a mechanism for reasoning about rules for goals, which are conditioned to some additional factors. In the simplest case, this can be done by adding such factors as new antecedents of the rules to be contextualised. But transformations may be problematic when complex reasoning patterns are considered. The framework of this paper is based on the following assumptions: Modalities: The system develops a constructive account of the modalities corresponding to mental states and obligations: rules are meant to devise the logical conditions for introducing them. Modalities may have a different logical behaviour. (Consider the special role played by belief rules, which here are not contextualised and permit to derive only unmodalised literals, whereas the other rule types allow for deriving modalised conclusions.) [7,9,5]. Conversions: Possible conversions of a modality into another can be accepted, as when the applicability of rule leading to derive, for example, OBLp (p is obligatory) may permit, under appropriate conditions, to obtain INTp (p is intended) [7,9]. Preferences: Preferences can be expressed in two ways: using standard DL superiority relation over rules and the operator ⊗. Operator ⊗ [8] applies to literals [5] as well as to rules, and captures the idea of violation. A ⊗-sequence such as α ⊗ β ⊗ γ
100
M. Dastani et al.
means that α is preferred, but if α is violated, then β is preferred; if β is violated, then the third choice is γ . Meta-rules: Meta-rules permit to reason about rules for deriving goals. This is the main device for contextualising the provability of goals and requires to introduce nested rules. We extend the language of Defeasible Logic with the modal operators BEL, INT, DES and OBL, and the non-classical connective ⊗. We divide the rules into meta-rules, and atomic rules. Atomic rules are in addition divided into rules for beliefs, desires, intentions, and obligations. For X ∈ {C, BEL, INT, DES, OBL}, where {BEL, INT, DES, OBL} is the set of modalities and C stands for contextual or meta-rules, we have that φ1 , . . . , φn →X ψ is a strict rule such that whenever the premises φ1 , . . . , φn are indisputable so is the conclusion ψ . φ1 , . . . , φn ⇒X ψ is a defeasible rule that can be defeated by contrary evidence. φ1 , . . . , φn ;X ψ is a defeater that is used to defeat some defeasible rules by producing evidence to the contrary. Definition 1 (Language). Let PROP be a set of propositional atoms and Lab be a set of labels. – The set of modal operators is MOD = {BEL, OBL, INT, DES}; – The set of literals is L = PROP ∪ {¬p|p ∈ PROP}. If q is a literal, ∼q denotes the complementary literal (if q is a positive literal p then ∼q is ¬p; and if q is ¬p, then ∼q is p); – The set of modal literals is MLit = {Xl, ¬Xl|l ∈ L, X ∈ {DES, INT, OBL}}; – The set of ⊗-expressions is PREF = {l1 ⊗ · · · ⊗ ln : n ≥ 1, {l1 , . . . , ln } ⊆ L}. n
We also write i=1 li for l1 ⊗ · · · ⊗ ln ∈ Pref. – The set of labeled atomic rules is Ruleatom = Ruleatom, s ∪ Ruleatom, d ∪ Ruleatom, dft , where for X ∈ MOD Ruleatom, s = {r : φ1 , . . . , φn →X ψ |r ∈ Lab, A(r) ⊆ L ∪ MLit, ψ ∈ L} Ruleatom, d = {r : φ1 , . . . , φn ⇒X ψ |r ∈ Lab, A(r) ⊆ L ∪ MLit, ψ ∈ Pref} Ruleatom, dft = {r : φ ;X ψ |r ∈ Lab, A(r) ⊆ L ∪ MLit, ψ ∈ Pref} – the set of labeled rules is Rule = Ruleatom ∪ {¬(r : φ1 , . . . , φn Y ψ )|(r : φ1 , . . . , φn Y ψ ) ∈ Ruleatom , ∈ {→, ⇒, ;},Y ∈ {DES, INT, OBL}} By convention, if r is a rule, ∼r denotes the complementary rule (if r : φ1 , . . . , φn X ψ then ∼r is ¬(r : φ1 , . . . , φn X ψ ); and if r : ¬(r : φ1 , . . . , φn X ψ ) then ∼r is r : φ1 , . . . , φn x ψ ). – the set of ⊗-rules is Q = {a1 ⊗ · · · ⊗ an|n ≥ 1, {a1 , . . . , an } ⊆ Rule}
Contextual Agent Deliberation in Defeasible Logic
101
– The set of labeled meta-rules is RuleC = RuleCs ∪ RuleCd ∪ RuleCdft , where, for X ∈ {DES, INT, OBL} RuleCs = {r : φ →C ψ |r ∈ Lab, φ ⊆ L ∪ MLit, ψ ∈ RuleX } RuleCd = {r : φ ⇒C ψ |r ∈ Lab, φ ⊆ L ∪ MLit, ψ ∈ Q} RuleCdft = {r : φ ;C ψ |r ∈ Lab, φ ⊆ L ∪ MLit, ψ ∈ Q} – We use some abbreviations, such as superscript for mental attitude or meta-rule, subscript for type of rule, and Rule[φ ] for rules whose consequent is φ , for example: RuleBEL = {r : φ1 , . . . , φn BEL ψ |(r : φ1 , . . . , φn BEL ψ ) ∈ Rule, ∈ {→, ⇒, ;}} Rules [ψ ] = {φ1 , . . . , φn →X ψ |{φ1 , . . . , φn } ⊆ L ∪ MLit, ψ ∈ L, X ∈ MOD} Other abbreviations are the following. We use r1 , . . . , rn to label (or name) rules, A(r) to denote the set {φ1 , . . . , φn } of antecedents of the rule r, and C(r) to denote the consequent ψ of the rule r. For some i, 1 ≤ i ≤ n, such that ci = q, R[ci = q] and rdX [ci = q] denote, respectively, the set of rules and a defeasible rule of type X with the head ⊗ni=1 ci such that ci = q. A defeasible agent theory consists of a set of facts or indisputable statements, a set of rules for beliefs, a set of meta-rules, a superiority relation > among rules saying when one rule may override the conclusion of another rule, and a conversion function c saying when a rule of one type can be used also as another type. Belief rules are the reasoning core of the agent. Rules for goals (desires, intentions, and obligations) are viewed in any theory as meta-rules with an empty antecedent and a consequent consisting of a ⊗-sequence of rules for goals. Definition 2 (Contextual Defeasible Agent Theory). A contextual defeasible agent theory is a structure D = (F, RBEL , RC , >, c) where F ⊆ L ∪ MLit is a finite set of facts, RBEL ⊆ RuleBEL , RC ⊆ RuleC , >⊆ (Rule × Rule) ∪ (RC × RC ), the superiority relation, is an acyclic binary relation over the set of rules; – c ⊆ MOD × MOD is a set of conversions.
– – – –
For readability reasons, we sometimes omit arrows for meta-rules r⇒C with the empty body. That is, a defeasible meta-rule ⇒C (p →INT q) is just represented as p →INT q. This extension of DL makes it possible to express ordered preferences over different options for contextualising rules for goals. In fact, we may have meta-rules such as the following: r : a ⇒C (r : b ⇒OBL c) ⊗ ¬(r : d ⇒INT f ⊗ g) Intuitively, meta-rule r states that, under the condition a, we should infer rule r stating that c is obligatory if b is the case; however, if this rule is violated (i.e., if, given b we obtain ¬c) then the second choice is to derive the negation of rule r , which would imply to intend f , as a first choice, or g as a second choice, if d is the case.
102
M. Dastani et al.
The following running example illustrates the contextual defeasible agent theory. Example 1. (RUNNING EXAMPLE). Frodo, our Tolkienian agent, intends to be entrusted by Elrond to be the bearer of the ring of power, a ring forged by the dark lord Sauron. Frodo has the task to bring the ring to Mordor, the realm of Sauron, and to destroy it by throwing it into the fires of Mount Doom. Given this task, if Frodo does not destroy the ring, he is obliged to leave the Middle Earth, while, if his primary intention is not to destroy it, but he does accomplish the task anyway, he will intend to go back to the Shire. If Frodo is a brave hobbit, rule r0 should hold, which states the intention to kill Sauron; however, if r0 is violated, then rule r3 should not hold. On the other hand, if Frodo is selfish, that he has the intention to be entrusted implies that he has also the intention to kill Elrond. However, if he has this last intention, he is not obliged destroy the ring if he is obliged to be the ring bearer. As facts of the theory, we know that Frodo intends to be entrusted by Elrond, that he is selfish and brave at the same time, and that he does not kill Sauron. F = {INTEntrusted, Selfish, Brave, ¬KillSauron} R = {r1 : OBLMordor ⇒OBL DestroyRing ⊗ LeaveMiddleEarth r2 : INTRingBearer ⇒OBL Mordor r3 : INTRingBearer ⇒INT ¬DestroyRing ⊗ BackToShire r4 : Entrusted ⇒OBL RingBearer r5 : Brave ⇒C (r0 : 0/ ⇒INT KillSauron) ⊗ ¬(r3 : INTRingBearer ⇒INT ¬DestroyRing ⊗ BackToShire) r6 : Selfish ⇒C (r7 : INTEntrusted →INT KillElrond) r8 : Selfish, INTKillElrond ⇒C (r9 : INTRingBearer ⇒OBL ¬DestroyRing)} >= {r3 > r1 , r5 > r8 } c = {c(OBL, INT)}
3 Reasoning about Contextual Deliberation Let X ∈ {C, BEL, DES, INT, OBL}. Proofs are sequences of literals, modal literals, and rules together with so-called proof tags +Δ , −Δ , +∂ and −∂ . Given a defeasible agent theory D, +ΔX q means that a conclusion q is provable in D using only facts and strict rules for X, −ΔX q means that it has been proved in D that q is not definitely provable in D, +∂X q means that q is defeasibly provable in D, and −∂X q means that it has been proved in D that q is not defeasibly provable in D. Before presenting proof procedures to derive specific tagged literals and rules in a contextual agent theory, we need to introduce some auxiliary notions. Definitions 3 and 4 are propaedeutic for Definition 5 (which defines the set of metarules supporting the derivation of a chosen rule) and Definition 6 (which defines the maximal-provable sets of rules that are provable in a theory). Definition 3 (Sub Rule). Let r ∈ Rule be an atomic rule and ∈ {→, ⇒, ;}. The set Sub(r) of sub-rules is defined as follows:
Contextual Agent Deliberation in Defeasible Logic
103
j – Sub(r) = {A(r) X ⊗i=1 ai |C(r) = ⊗ni=1 ai , j ≤ n}, if r is atomic j – Sub(r) = {¬(A(r) X ⊗i=1 ai )|C(r) = ⊗ni=1 ai , j ≤ n}, otherwise
E.g., given r : (a →INT b ⊗ c), Sub(r) = {a →INT b, a →INT b ⊗ c}. Definition 4 (Modal Free Rule). Given an atomic rule r, the modal free rule L(r) of r is obtained by removing all modal operators in A(r). For example, given r : INTa →INT b, L(r) is r : a →INT b. Definition 5 (Rule-Supporting Rules). Let D be a contextual agent theory. The set RC rX of supporting rules in RC for a non-nested rule rX ∈ Rule is: – if rX ∈ Ruleatom and ∀a ∈ A(r) : a = Xb ∈ MLit, C X C X C Y R r = R [ci = L(s )] R [ci = s ] ∪ sX ∈Sub(rX )
– otherwise
RC rX =
Y :c(Y,X)
RC [ci = sX ]
∀sX ∈Sub(rX )
For example, a meta-rule ⇒C (a ⇒INT b ⊗ c) ⊗ (a ⇒INT d) supports the following rules: (a ⇒INT b), (a ⇒INT b ⊗ c), and (a ⇒INT d). Definition 6 (Maximal Provable-Rule-Sets). Let D be a contextual agent theory. The maximal provable-rule-sets of non-nested rules that are possibly provable in D is, for X ∈ {DES, INT, OBL}, n X C ci , r ∈ R ∪ RP = Sub(ci)|C(r) =
i=1 X n C Y Sub(L(c i ))|∀Y such that sc(Y, X),C(r) = ⊗i=1 ci , r ∈ R , and ∀a ∈ A(r) : a = Xb ∈ MLit
RPBEL ={Sub(r)|r ∈ RBEL }. Since we want to derive rules for goals, this requires defining when two rules are incompatible. In this regard, notice that defeasible rules and defeaters for goals may have ⊗-expressions as their consequents, which is something we do not have for strict rules. Accordingly, the notion of incompatibility has to take also into account when two ⊗expressions occur in the the heads of two rules. Definition 7. Two non-nested rules r and r are incompatible iff r is an incompatible atomic rule of r or r is an incompatible negative rule of r. rule of r iff r and r are atomic rules and 1. r is an incompatible atomic n – A(r) = A(r ), C(r) = i=1 ai and – C(r ) = m i=1 bi , such that ∃ j, 1 ≤ j ≤ n, m, a j = ∼b j and, ∀ j ≤ j, a j = b j .
104
M. Dastani et al.
2. r is an incompatible negative rule of r iff either r or r is not an atomic rule and – A(r) = A(r ), C(r) = ni=1 ai and – C(r ) = m i=1 bi , such that N = min{n, m}, ∀ j ≤ N, a j = b j . Definition 8. Let D be a contextual agent theory and r a non-nested rule. The set of all possible incompatible rules for rX is:
IC(rX ) = r |r ∈ RPX , r is incompatible with rX Definitions 9 and 10 define, respectively, when a literal or a rule is provable (and nonprovable: rejected), and when a rule is applicable (and non-applicable: discarded). Both notions are used in the proof procedures presented in the remainder. Definition 9 (Provable). Let # ∈ {Δ , ∂ }, P = (P(1), . . . , P(n)) be a proof in a contextual agent theory D, and X ∈ {DES, INT, OBL}. A literal q ∈ L or a rule r ∈ Rule are #-provable in P if there is an initial sequence P(1), . . . , P(m) of P such that either 1. 2. 3. 4.
q is a literal and P(m) = +#BELq or q is a modal literal X p and P(m) = +#X p or q is a modal literal ¬X p and P(m) = −#X p or rX is a rule in RPX and P(m) = +#C rX ;
A literal q ∈ L or a rule r ∈ Rule are #-rejected in P if there is an initial sequence P(1), . . . , P(m) of P such that either 1. 2. 3. 4.
q is a literal and P(m) = −#BELq or q is a modal literal X p and P(m) = −#X p or q is a modal literal ¬X p and P(m) = +#X p or rX is a rule in RPX and P(m) = −#C rX .
Definition 10. Let D be a contextual agent theory. Applicable rules and discarded rules are defined as follows: 1. A rule r ∈ RBEL ∪ RC is applicable iff ∀a ∈ A(r): – if a ∈ L then +∂BELa ∈ P(1..n), and – if a = Xb ∈ MLit then +∂X a ∈ P(1..n). 2. A rule r ∈ R[ci = q] is applicable in the condition for ±∂X iff – r ∈ RXatom and ∀a ∈ A(r): if a ∈ L then +∂BEL a ∈ P(1..n), and if a = Zb ∈ MLit then +∂Z a ∈ P(1..n); or – r ∈ RYatom and c(Y, X) ∈ c and ∀a ∈ A(r): +∂X a ∈ P(1..n). 3. A rule r is discarded in the condition for ±∂X iff either: – if r ∈ RBEL ∪ RC ∪ RX then either ∃a ∈ A(r) : −∂BEL a ∈ P(1..n) or ∃Xb ∈ A(R), Xb ∈ MLit and −∂X b ∈ P(1..n); or – if r ∈ RY , then ∃a ∈ A(r) : −∂X a ∈ P(1..n). Notice that the notion of applicability needs to take conversions into account.
Contextual Agent Deliberation in Defeasible Logic
105
Remark 1. Conversions affect applicability of rules. In many cases we want that this possibility can be admitted: +ΔINT GoToRome, GoToRome →BEL GoToItaly |∼ +ΔINT GoToItaly +∂INT GoToRome, GoToRome ⇒OBL VisitVatican |∼ +∂INT VisitVatican The first rule states that the agent believes that going to Rome strictly implies going to Italy; if we can derive that the agent has the intention to visit Rome, we have reasons to say that a rational agent may have the intention to visit Italy. The second rule says that visiting Rome defeasibly implies the obligation to visit Vatican City. With norm-complying agents, agent’s intention to visit Rome rationally implies to having the intention to go to Vatican City. Example 2. Suppose we allow a deontic rule to be converted into a rule for intention, i.e. c(OBL, INT). Consider rule r : a, b ⇒OBL p: if ∂INT a and ∂INT b, then r is applicable in the proof condition for +∂INT . Before providing proof procedures to derive rules, let us introduce specific proof tags for this purpose. Remember that denotes either →, ⇒ or ; to simplify our presentation. ±ΔC rX means that rule r ∈ RX is (is not) definitely provable using meta-rules; ±∂C rX means that rule r ∈ RX is (is not) defeasibly provable using meta-rules. In general, ±ΔCX and ±∂CX mean, respectively, definitive (non-)provability of rules for X, and defeasible (non-)provability of rules for X. Let us see proof procedures to derive rules. In this perspective, however, we have to be careful, as we can distinguish between strict and defeasible derivations of nonnested strict and defeasible rules. Given a contextual agent theory D, a non-nested rule r is strictly provable in D when it is strictly derived using a meta-rule such as a →C r. A rule r is defeasibly provable in D when it is defeasibly derived using a meta-rule such as a →C r and a ⇒C r. When a strict atomic rule a →INT b is defeasibly derived, it acts as a defeasible rule a ⇒INT b. Proof procedures for the strict derivation of atomic rules in a contextual defeasible agent theory D are as follows: +ΔCX : If P(i + 1) = +ΔC rX then (1) X = BEL and rX ∈ RBEL or (2) ∃s ∈ RCs rX ∀a ∈ A(s) a is Δ -provable. −ΔCX : If P(i + 1) = −ΔC rX then (1) X = BEL or rX ∈ RBEL and (2) ∀s ∈ RCs rX ∃a ∈ A(r) : a is Δ -rejected. Strict derivations of rules are based on the following intuition. If the rule r we want to derive is for belief, r must be in the set of belief rules of the theory. Otherwise (for the other rule types), r must be proved using a strict meta-rule whose antecedents are strictly provable. Defeasible derivations of rules are based on the following procedures.
106
M. Dastani et al.
+∂CX : If P(n + 1) = +∂C rX , then (1) +ΔC rX ∈ P(1..n), or (2) (1) ∀r ∈ IC(rX ), ∀r ∈ RCs r , r is discarded and (2) ∃t ∈ RC ci = rX such that (1) ∀i < i, ci is applicable, (2) ∀i < i, C(ci ) = nk=1 bk , such that ∀k : +∂BEL∼bk ∈ P(1..n), (3) t is applicable, and (3) ∀r ∈ IC(rX ), ∀s ∈ RC di = r (1) if ∀i < i, di is applicable, C(di ) = nk=1 ak s.t. ∀k : +∂BEL ∼ak ∈ P(1..n), then (1) s is discarded, or (2) ∃z ∈ RC pi = r : r ∈ IC(C(s)) s.t. ∀i < i, pi is applicable, and C(pi ) = nk=1 dk s.t. ∀k : +∂BEL ∼dk ∈ P(1..n) and z is applicable and z > s. −∂CX : If P(n + 1) = −∂C rX , then (1) −ΔC rX ∈ P(1..n), and (2) (1) ∃r ∈ RCs r such that r ∈ IC(rX ), r is applicable or (2) ∀t ∈ RC ci = rX (1) ∃i < i such that ci is discarded, or (2) ∃i < i such C(ci ) = nk=1 bk and ∃k : −∂BEL ∼bk ∈ P(1..n), (3) t is discarded, or (3) ∃s ∈ RC di = r such that ∀r ∈ IC(rX ), such that (1) ∀i < i, di is applicable, C(di ) = nk=1 ak s.t. ∀k : +∂BEL ∼ak ∈ P(1..n), and (1) s is applicable, and (2) ∀z ∈ RC pi = r such that r ∈ IC(C(s)) ∃i < i, pi is discarded or C(pi ) = nk=1 dk s.t. ∃k : −∂BEL ∼dk ∈ P(1..n) or z is discarded or z > s.
Remark 2. The defeasible proof of a rule runs in three phases. We have to find an argument in favour of the rule we want to prove. Second, all counter-arguments are examined (rules for the opposite conclusion). Third, all the counter-arguments have to be rebutted (the counter-argument is weaker than the pro-argument) or undercut (some of the premises of the counter-argument are not provable). Let us exemplify positive proof conditions (+∂CX ) step by step. Suppose we want to derive r : INTa →INT b, namely, that X = INT and =→. We have the following options. Condition (1): r is definitely provable; or, Condition (2): We use a strict or defeasible meta-rule to derive r. This must exclude, as a precondition, that any rule, which is incompatible with r, is definitely supported: (condition 2.1). That is, rules such as r : ¬(INTa →INT b)
r : INTa →INT ¬b
r : a →OBL ¬b
should not be supported, if we have that r (the rule we want to prove) is applicable, +∂INT a and c(OBL, INT), namely, if we may convert a rule for obligation into one for intention. In fact, if we have this conversion, r behaves like a rule for intention. With this done, condition (2.2) states that there should exist a meta-rule such as t : d ⇒C (w : p ⇒OBL q) ⊗ (r : INTa →INT b) such that t is applicable, +∂BELd, and the first choice, rule w, is violated, namely that +∂BEL p and +∂BEL ¬q. But this fact must exclude that any meta-rule s supporting an incompatible conclusion against r is applicable (see condition 2.3.1.1). Alternatively, if
Contextual Agent Deliberation in Defeasible Logic
107
s is applicable, we have to verify that there exists a meta-rule z supporting r such that z is applicable and is stronger that s (see condition 2.3.1.2). Notice that when we say that that a meta-rule supports a rule, we take into account that meta-rules may have ⊗-rules in their consequents. For example, if s is as follows s : d ⇒C (w : b →DES c) ⊗ r : ¬(INTa →INT b), but we prove −∂BEL¬c, this means that w cannot be violated and so s cannot be used to attack the derivation of r: in this case, using s, we could only prove rule w. Given the above proof conditions for deriving rules, the following are the procedures for proving literals. Notice that each time a rule r is used and applied, we are required to check that r is provable. +ΔX : If P(i + 1) = +ΔX q then (1) Xq ∈ F, or q ∈ F if X = BEL, or (2) ∃r ∈ RuleXs [q] : +ΔC r and ∀a ∈ A(r) a is Δ -provable or (3) ∃r ∈ RuleYs [q] : +ΔC r, ∀a ∈ A(r) a is Δ -provable and c(Y, X). −ΔX : If P(i + 1) = −ΔX q then (1) Xq ∈ / F, or q ∈ / F if X = BEL, and (2) ∀r ∈ RuleXs [q] : −ΔC r or ∃a ∈ A(r) : a is Δ -rejected and (3) ∀r ∈ RuleYs [q] : −ΔC r, or if c(Y, X) then ∃a ∈ A(r) a is Δ -rejected. +∂X : If P(n + 1) = +∂X q then (1)+ΔX q ∈ P(1..n) or (2) (1) −ΔX ∼q ∈ P(1..n) and (2) ∃r ∈ Rulesd [ci = q] such that +∂C r, r is applicable, and ∀i < i, +∂BEL ∼ci ∈ P(1..n); and (3) ∀s ∈ Rule[c j = ∼q], either −∂C s, or s is discarded, or ∃ j < j such that −∂BEL∼c j ∈ P(1..n), or (1) ∃t ∈ Rule[ck = q] such that +∂C t, t is applicable and ∀k < k, +∂BEL ∼ck ∈ P(1..n) and t > s. −∂X : If P(n + 1) = −∂X q then (1)−ΔX q ∈ P(1..n) and either (2) (1) +ΔX ∼q ∈ P(1..n) or (2) ∀r ∈ Rulesd [ci = q], either −∂C r, or r is discarded or ∃i < i such that −∂BEL∼ci ∈ P(1..n); or (3) ∃s ∈ Rule[c j = ∼q] such that +∂C s, s is applicable and ∀ j < j, +∂BEL ∼c j ∈ P(1..n), and (1) ∀t ∈ Rule[ck = q] either −∂C t, or t is discarded, or ∃k < k such that −∂BEL∼ck ∈ P(1..n) or t > s. Example 3. (RUNNING EXAMPLE , CONTINUED). The fact that Frodo has the intention to be entrusted by Elrond makes it possible to derive +ΔINT Entrusted. Since we have the conversion c(OBL, INT), this would make both r2 and r3 applicable. However, r5 and r6 , too, are applicable. Rule r6 permits to derive r7 , which is applicable, as Frodo
108
M. Dastani et al.
has the intention to be entrusted. This allows in turn for the derivation of the intention to kill Elrond, which makes r8 applicable. Here we have a conflict between r5 and r8 , but the former is stronger. On the other hand, r5 states, as a first choice, that Frodo has the intention to kill Sauron. But this intention is violated, as ¬KillSauron is a fact, which makes it possible to derive +ΔBEL ¬KillSauron and so +∂BEL¬KillSauron. Thus we have to derive the second choice, namely, the negation of r3 . In this way, even if r3 is stronger than r1 , the applicability of meta-rule r5 makes r3 inapplicable.
4 Context-Detection: A Further Example Context can play as a disambiguating function in agent communication. For example, when an agent A receives a command “On” from another agent B, the meaning of the message usually depends on the common context between the two agents. Without context information, the command “On” is too ambiguous because “On” could mean many things, for example, turn on a water tap, turn on a light, turn on a projector, or run a weekly meeting presentation. To show how contexts can be detected and decisions can be made using contextual information within our reasoning model, we consider a simple theory representing an office assistant that can help office workers to control their modernized office environment for non-interrupted work flow: F = {MRoom, Monday, Morning, onPro jector} RBEL = {r1 : MRoom, Monday, Morning ⇒BEL CW Meeting r2 : MRoom, Monday, Morning ⇒BEL ¬CDMeeting r3 : MRoom, Morning ⇒BEL CDMeeting r4 : onPro jector →BEL ¬trunOnPro jector} R = {r5 : CW Meeting ⇒C (MessageOn ⇒INT turnOnPro jector ⊗ C
openW MPresentation ⊗ openDMPresentation) r6 : CDMeeting ⇒C (MessageOn ⇒INT turnOnPro jector ⊗ openDMPresentation ⊗ openWMPresentation)} Herein, rules r1 and r2 say that if the assistant agent is in a meeting room (MRoom) and it is Monday morning, then usually the context is that of a weekly meeting (CW Meeting) and not of a daily meeting (CDMeeting). However, rule r3 says that if the assistant agent is in a meeting room in the morning, then the context is usually that of a daily meeting. That is, r1 , r2 , and r3 are used to detect the context. Once the context is determined, the assistant agent can properly process the command “On”. For instance, r5 says that if it is a weekly meeting, enable the following rule: MessageOn ⇒INT trunOnPro jector ⊗ openWMPresentation ⊗ openDMPresentation meaning that, if the agent receives a command “On”, then the agent usually should form the intention to turn on the data projector, but if it cannot be turned on (because it is already turned on), then it should try to open a weekly meeting presentation file.
Contextual Agent Deliberation in Defeasible Logic
109
However, for some reason, if the weekly meeting presentation cannot be run (maybe it is just a daily meeting), then it should try to open a daily meeting presentation file. On the other hand, r6 says that if it is a daily meeting, enable the following rule: Message On ⇒INT trunOnPro jector ⊗ openDMPresentation ⊗ openWMPresentation This rule sasys that, in this context, opening a daily meeting presentation file (openDMPresentation) has higher priority than opening a weekly meeting presentation file (W MPresentation). With the given theory, the office agent will conclude +∂INT openWMPresentation if it receives Message On because it is a weekly meeting and a projector is already turned on. This example clearly illustrates how contextual information is naturally represented within our reasoning model. That is, contextual information can be used to enable certain rules and to change priority between deliberations.
5 Summary We extended Defeasible Logic to deal with the contextual deliberation process of cognitive agents. First, we introduce meta-rules to reason with rules. Meta-rules are rules that have, as a consequent, rules to derive goals (obligations, intentions and desires): in other words, meta-rules include nested rules. Second, we introduce explicit preferences among rules. They deal with complex structures where nested rules can be involved to capture scenarios where rules are violated. Further research are the development of a methodology to use the language, and a formal analysis of the logic.
References 1. Antoniou, G., Billington, D., Governatori, G., Mahe, M.J.: Representation results for defeasible logic. ACM Transactions on Computational Logic 2(2), 255–287 (2001) 2. Antoniou, G., Billington, D., Governatori, G., Maher, M.J.: Embedding Defeasible Logic in Logic Programming. Theory and Practice of Logic Programming 6(6), 703–735 (2006) 3. Broersen, J., Dastani, M., Hulstijn, J., van der Torre, L.: Goal generation in the BOID architecture. Cog. Sc. Quart. 2(3-4), 428–447 (2002) 4. Cohen, P.R., Levesque, H.J.: Intention is choice with commitment. Artif. Intell. 42(2-3), 213–261 (1990) 5. Dastani, M., Governatori, G., Rotolo, A., van der Torre, L.: Programming cognitive agents in defeasible logic. In: Sutcliffe, G., Voronkov, A. (eds.) LPAR 2005. LNCS (LNAI), vol. 3835, pp. 621–636. Springer, Heidelberg (2005) 6. Governatori, G., Padmanabhan, V., Rotolo, A.: Rule-based agents in temporalised defeasible logic. In: Yang, Q., Webb, G. (eds.) PRICAI 2006. LNCS (LNAI), vol. 4099, pp. 31–40. Springer, Heidelberg (2006) 7. Governatori, G., Rotolo, A.: Defeasible logic: Agency, intention and obligation. In: Lomuscio, A., Nute, D. (eds.) DEON 2004. LNCS, vol. 3065, pp. 114–128. Springer, Heidelberg (2004) 8. Governatori, G., Rotolo, A.: Logic of violations: A Gentzen system for reasoning with contrary-to-duty obligations. Australasian Journal of Logic 4, 193–215 (2006) 9. Governatori, G., Rotolo, A., Padmanabhan, V.: The cost of social agents. In: Proc. AAMAS 2006, pp. 513–520 (2006) 10. Rao, A.S., Georgeff, M.P.: Decision procedures for bdi logics. J. Log. Comput. 8(3), 293–342 (1998) 11. Song, I., Governatori, G.: Nested rules in defeasible logic. In: Adi, A., Stoutenburg, S., Tabet, S. (eds.) RuleML 2005. LNCS, vol. 3791, pp. 204–208. Springer, Heidelberg (2005)
Real-Time Moving Target Search Cagatay Undeger1 and Faruk Polat2 1
Savunma Teknolojileri Muhendislik ve Tic.A.S., 06510 Ankara, Turkey 2 Middle East Technical University, 06531 Ankara, Turkey [email protected], [email protected]
Abstract. In this paper, we propose a real-time moving target search algorithm for dynamic and partially observable environments, modeled as grid world. The proposed algorithm, Real-time Moving Target Evaluation Search (MTES), is able to detect the closed directions around the agent, and determine the best direction that avoids the nearby obstacles, leading to a moving target which is assumed to be escaping almost optimally. We compared our proposal with Moving Target Search (MTS) and observed a significant improvement in the solution paths. Furthermore, we also tested our algorithm against A* in order to report quality of our solutions. Keywords: Path Planning, Real-Time Search, Moving Target Search.
1 Introduction Pursuing a moving target is one of the most challenging problems in areas such as robotics, computer games, etc. Off-line and incremental path planning algorithms are not able to handle moving targets in real-time, and most of the on-line search algorithms are specifically designed for partially observable environments with static targets. The most well known algorithm for moving targets is Moving Target Search (MTS) [2], which maintains a heuristic table that contains estimated costs of paths between every pair of coordinates. In this paper, we propose a new moving target search algorithm, Real-Time Moving Target Evaluation Search (MTES), which is build on Real-Time Target Evaluation Search [9] developed for partially observable environments with static targets. MTES is able to estimate the distance to the target considering the intervening obstacles and discard some non-promising alternative moving directions in real-time in order to guide the agent to a moving target. The method sends virtual (non-physical) rays away from the agent in four directions, and determines the obstacles that the rays hit. For each such obstacle, we extract its border and determine the best direction that avoids the obstacle if the target is blocked by the obstacle. Hence, we have a number of directions each avoiding an obstacle hit by a ray. Then by using these directions and a resolution mechanism, a single moving direction is determined. To show the performance of our algorithm, we compared MTES with MTS-c, MTS-d and A*. For the experiments, we randomly generated grids of different types, and developed a successful prey algorithm (Prey-A*) in order to challenge the algorithms used in the experiments. The results showed that MTES produces near optimal solutions, and outperforms MTS-c and MTS-d significantly. A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 110–121, 2009. c Springer-Verlag Berlin Heidelberg 2009
Real-Time Moving Target Search
111
The related work is given in Section 2. In section 3, MTES is described in details. In Section 4, the performance analysis of MTES is presented. Finally Section 5 is the conclusion.
2 Related Work Off-line path planning algorithms [6] are hard to use for large dynamic environments and for moving targets because of their time requirements. One way to make these algorithms more efficient is to change them from off-line to incremental [7,4] in order to avoid re-planning from scratch. Although these algorithms work fine with partially observable environments, they are sometimes not efficient enough and usually not capable of handling moving targets. There are also a number of on-line approaches [5,3,8,9,2]. As a matter of fact, only few of these on-line algorithms can be adapted against a moving target. Moving Target Search (MTS) [2] is a well known real-time algorithm for pursuing a moving target, which is built on Learning Real-Time A*. The algorithm maintains a table of heuristic values, presenting the function h(x, y) for all pairs of locations x and y in the environment, where x is the location of the agent and y is the location of the target. The original MTS is a poor algorithm in practice because when the target moves, the learning process has to start all over again. Therefore, two MTS extensions namely Commitment to Goal (MTS-c) and Deliberation (MTS-d) [2] are proposed to improve the solution quality. In order to use the learned table values more effectively, MTS-c ignores some of the target’s moves, and MTS-d performs an off-line search while in a heuristic depression. When we look at the prey algorithms, we usually cannot see very successful studies. Mostly the focus is on the predators, and the prey algorithms commonly use hybrid techniques mixing the reactive strategies [1,2]. Since these strategies are not challenging enough for our algorithms, we developed an off-line strategy, which is slow but more powerful with respect to its escape capability.
3 Moving Target Search 3.1 Problem Description In our study, the environment is a grid world, where any grid cell can either be free or obstacle. There is a single agent (predator) that aims to reach a static or moving target (prey). Both are randomly located far from each other in non-obstacle grid cells. The predator is expected to reach the prey from a short path avoiding obstacles in real-time. We assume that the predator knows the prey’s location all the time, but has limited perception, and is only able to sense the obstacles around him within a square region centered at the agent location. The size of the square is (2v + 1)x(2v + 1), where v is the vision range. We used the term infinite vision to emphasize that the agent has unlimited sensing capability and knows the entire grid world before the search starts. The unknown parts of the grid world is assumed to be free of obstacle by the agent, until it is explored. Therefore, when we say an obstacle, we refer to the known part of
112
C. Undeger and F. Polat
that obstacle. The agent can only perform four actions in each step; moving to a free neighbor cell in north, south, east or west direction. The prey has unlimited perception and knows all the grid world and the location of the predator all the time. The search continues until the predator reaches the prey. 3.2 MTES Algorithm MTES makes use of a heuristic, Real-Time Target Evaluation (RTTE-h), which analyzes obstacles and proposes a moving direction that avoids these obstacles and leads to the target through shorter paths. To do this, RTTE-h geometrically analyzes the obstacles nearby, tries to estimate the lengths of paths around the obstacles to reach the target, and proposes a moving direction. RTTE-h works in continuous space to identify the moving direction, which is then mapped to one of the actual moving directions (north, south, east and west). MTES repeats steps in Algorithm 1 until reaching the target or detecting that target is inaccessible. In the first step, MTES calls RTTE-h heuristic function, which returns a moving direction and the utilities of neighbor cells according to that proposed direction. Next, MTES selects one of the neighbor cells on open directions with the minimum visit count, which stores the number of visits to the cell. If there exists more than one cell having the minimum visit count, the one with the maximum utility is selected. If utilities are also the same, then one of them is selected randomly. After the move is performed, the visit count of the previous cell is incremented and the cell is inserted into the history. The set of previously visited cells forms the history of the agent. History cells are treated as obstacles. Therefore, if the agent discovers a new obstacle during the exploration and realizes that the target became inaccessible due to history cells, the agent clears the history to be able to backtrack. Algorithm 1. An Iteration of MTES Algorithm 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14:
Call RTTE-h to compute the proposed direction and the utilities of neighbor cells. if a direction is proposed by RTTE-h then Select the neighbor cell with the highest utility from the set of non-obstacle neighbors with the smallest visit count. Move to the selected direction. Increment the visit count of previous cell by one. Insert the previous cell into the history. else if History is not empty then Clear all the History. Jump to 1 else Destination is unreachable, stop the search with failure. end if end if
In moving target search problem, the target may sometimes pass through the cells the agent previously walked through. In such a case, there is a risk that the history blocks the agent to reach the target since history cells are assumed to be obstacles and may close some gateways required to return back. If this situation occurs at some point, the agent will surely be able to detect this at the end, and clear the history, opening all the closed gateways. Therefore, the algorithm is capable of searching moving targets without any
Real-Time Moving Target Search
113
Fig. 1. Sending rays to split moving directions
additions. As a matter of fact, the only drawback of the history is not the possibility that it can block the way to the target entirely, but it can sometimes prevent the agent to reach the target through shorter paths by just closing some of the shortcuts. To reduce the performance problems of this side effect, the following procedure is applied. Assuming that (x1 , y1 ) and (x2 , y2 ) are the previous and newly observed locations of the target, respectively, and R is the set of cells the target could have visited in going from (x1 , y1 ) to (x2 , y2 ), the algorithm clears the history along with visit counts when any cell in set R appears in history or has non-zero visit count. In the algorithm, R can be determined in several ways depending on the required accuracy. The smallest set has to contain at least the newly observed location of the target, (x2 , y2 ). One can choose to ignore some of the set members and only use (x2 , y2 ) to keep the algorithm simple. The RTTE-h heuristic method in Algorithm 2 propagates four diagonal rays away from the agent location (line 2 in Algorithm 2) to split north, south, east and west moving directions as shown in Fig 1. The rays move outwards from the agent until they hit an obstacle or maximum ray distance is achieved. Four rays split the area around the agent into four regions. A region is said to be closed if the target is inaccessible from any cell in that region. If all the regions are closed, the target is unreachable from the current location. To detect closed regions, the boundary of the obstacle is extracted (line 4) and analyzed (line 5). Next, the obstacle border is re-traced from both left and right sides to determine geometric features of the obstacle (line 6). These features are evaluated and a moving direction to avoid the obstacle is identified (line 7). After all the obstacles are evaluated, results are merged to propose a final moving direction (line 9). The obstacle border extraction and closed direction detection phases of the algorithm use the same methods presented in [8,9]. The geometric features of an obstacle are determined in two phases: left analysis and right analysis (line 6 in Algorithm 2). In left/right analysis, the known border of the obstacle is traced edge by edge towards the left/right sides starting from the hit point, making a complete tour around the obstacle border. During this process, several geometric features of the obstacle are extracted, which are described in Definitions 1 to 8 and illustrated in Fig. 2. Definition 1 (Outer left most direction). Relative to the ray direction, the largest cumulative angle is found during the left tour on the border vertices. In each step of the trace, we move from one edge vertex to another on the border. The angle between the
Mark all the moving directions as open. Propagate four diagonal rays. for each ray hitting an obstacle do Extract the border of the obstacle by tracing the edges from left side until making a complete tour around the obstacle. Detect closed directions. Extract geometric features of the obstacle. Evaluate the results and determine a direction to avoid the obstacle. end for Merge individual results, propose a direction to move, and compute utilities of neighbor cells.
two lines (TWLNS) starting from the agent location and passing through these two following vertices is added to the cumulative angle computed so far. Note that the added amount can be positive or negative depending on whether we move in counter-clockwise (ccw) or clockwise (cw) order, respectively. This trace (including the trace for the other geometric features) continues until the sum of the largest cumulative angle and the absolute value of smallest cumulative angle is greater than or equal to 360. The largest cumulative angle before the last step of trace is used as the outer left most direction. Definition 2 (Inner left most direction). The direction with the largest cumulative angle encountered during the left tour until reaching the first edge vertex where the angle increment is negative and the target lies between TWLNS. If such a situation is not encountered, the direction is assumed to be 0 + ε , where ε is a very small number (e.g., 0.01). Definition 3 (Inside of left). True if the target is inside the polygon whose vertices starts at agent’s location, jumps to outer left most point, follows the border of the obstacle to the right and ends at the hit point of the ray. Definition 4 (Inside of inner left). True if the target is inside the polygon that starts at agent’s location, jumps to the inner left most point, follows the border of the obstacle to the right and ends at the hit point of the ray. Definition 5 (Behind of left). True if the target is in the region obtained by sweeping the angle from the ray direction to the outer left most direction in ccw order and the target is not inside of left. Definition 6 (Outer-left-zero angle blocking). True if target is in the region obtained by sweeping the angle from the ray direction to the outer left most direction in ccw order. Definition 7 (Inner-left-zero angle blocking). True if target is in the region obtained by sweeping the angle from the ray direction to the inner left most direction in ccw order. In right analysis, the border of the obstacle is traced towards the right side and the same geometric properties listed above but now symmetric ones are identified. In the right analysis, additionally the following feature is extracted:
Real-Time Moving Target Search
115
Fig. 2. Geometric features of obstacles
Definition 8 (Left alternative point). The last vertex in the outer left most direction encountered during the right tour until the outer right most direction is determined (see Fig. 2). In obstacle evaluation step (line 7 in Algorithm 2), if an obstacle blocks line of sight from agent to target, we determine a moving direction that avoids the obstacle meanwhile reaching the target through a shorter path. Method is given in Algorithm 3, which requires path length estimations given in Definitions 9 to 11. Definition 9 (dle f t ). The approximated length of the path which starts from the agent location, jumps to the outer left most point, and then follows the path determined by Algorithm 4 (see Fig. 3). Definition 10 (dle f t.alter ). The approximated length of the path which starts from the agent location, jumps to the outer right most point, and then to the outer left most point, and finally follows the path determined by Algorithm 4 (see Fig. 3). Definition 11 (dle f t.inner ). The approximated length of the path passing through the agent location, the inner left most point, and the target (see Fig. 3).
if (behind of left and not inside of right) or (behind of right and not inside of left) then if outer left most angle + outer right most angle ≥ 360 then if distance from agent to outer left most point is smaller than distance from agent to left alternative point then Assign estimated distance as min(dle f t , dright.alter ) and propose outer left most direction as moving direction else Assign estimated distance as min(dle f t.alter , dright ) and propose outer right most direction as moving direction end if else if dle f t < dright then Assign estimated distance as dle f t and propose outer left most direction as moving direction else Assign estimated distance as dright and propose outer right most direction as moving direction end if end if Mark obstacle as blocking the target else if behind of left then if Target direction angle = 0 and outer-right-zero angle blocking then Assign estimated distance as dle f t and propose outer left most direction as moving direction else Assign estimated distance as dright.inner and propose inner right most direction as moving direction end if Mark obstacle as blocking the target else if behind of right then if Target direction angle = 0 and outer-left-zero angle blocking then Assign estimated distance as dright and propose outer right most direction as moving direction else Assign estimated distance as dle f t.inner and propose inner left most direction as moving direction end if Mark obstacle as blocking the target else if (inside of left and not inside of right) and (inner-left-zero angle blocking and not inside of inner left) then Assign estimated distance as dle f t.inner and propose inner left most direction as moving direction Mark obstacle as blocking the target else if (inside of right and not inside of left) and (inner-right-zero angle blocking and not inside of inner right) then Assign estimated distance as dright.inner and propose inner right most direction as moving direction Mark obstacle as blocking the target end if end if
Fig. 3. Examples of dle f t (left), dle f t.alter (middle) and dle f t.inner (right)
Algorithm 4 is internally used in computations of dle f t and dle f t.alter , and the subfunction isoutwardsfacing is called for detecting if a border segment, whose both ends touch the line passing through the outer left most point and the target point, is outwards facing ( see Fig. 4 ). The estimated target distances over right side of the obstacle are
Real-Time Moving Target Search
117
Algorithm 4. Path length estimation Require: t : target point Require: s : outer left most point Require: n : the nearest point to the target Require: + : next border point (left of) Require: − : previous border point (right of) Require: insert(p) : inserts a point to the estimated path Require: clasi f y(p1 , p2 , p3 ) : if the edge formed by the points p1 , p2 , p3 does a left turn then returns true, else returns false Require: isoutwards f acing(side, p1 , p2 ) : see Algorithm 5 1: let prev = s 2: let prevle f t = true 3: insert(s) 4: for each border point v between s+ and n do 5: if v = n then 6: if isoutwards f acing(prevle f t, prev,t) then 7: insert(all border points between prev+ and v) 8: end if 9: insert(t) 10: return length of estimated path 11: end if 12: let vle f t = not clasi f y(s,t, v) 13: if prevle f t = vle f t then 14: let z = intersection point of lines (s,t) and (v−, v) 15: if not isoutwards f acing(prevle f t, prev, z) and z is between prev and t then 16: insert(t) 17: return length of estimated path 18: end if 19: if isoutwards f acing(prevle f t, prev, z) then 20: insert(all border points between prev+ and v) 21: else 22: insert(v) 23: end if 24: let prev = v 25: let prevle f t = vle f t 26: end if 27: end for
Algorithm 5. The function isoutwards f acing(side, p1, p2 ) Require: t : target point Require: s : outer left most point Require: len(n1 , n2 ) : returns distance between points n1 and n2 Require: positive(m) : if len(m,t) ≤ len(s,t) or len(m,t) ≤ len(s, m) then returns true, else returns false Require: slen(m) : if positive(m) then returns +len(s, m) else returns −len(s, m) 1: if (side and slen(p1 ) < slen(p2 )) or (not side and slen(p1 ) > slen(p2 )) then 2: return true 3: else 4: return false 5: end if
similar to those over left side of the obstacle, and computed symmetrically (the terms left and right are interchanged). In the merging phase (line 9 in Algorithm 2), the evaluation results for all obstacles are used to determine a single moving direction [9]. This proposed direction will be passed to MTES algorithm (Algorithm 1) for final decision. The complexity of MTES is O(w.h) per step, where w is the width and h is the height of the grid world. Since increasing the grid size decreases the efficiency, a search depth (d) can be introduced in order to limit the worst case complexity. A search depth is a rectangular area of size (2d + 1)x(2d + 1) centered at agent location, which makes the
118
C. Undeger and F. Polat
Fig. 4. Exemplified path length estimation, and outwards/inwards facing segments
Fig. 5. Average of path length results of maze grids for increasing vision ranges
algorithm treat the cells beyond the rectangle as non-obstacle. With this limitation, the complexity becomes O(d 2 ).
4 Performance Analysis In this section, we present our experimental results on MTS-c, MTS-d, MTES and A*. As being an o-line algorithm, we executed A* in each step from scratch. For the test runs, we used 9 randomly generated sample grids of size 150x150. Six of them were the maze grids, and three of them were the U-type grids [8,9]. For each grid, we produced 15 different randomly generated predator-prey location pairs, and made all the algorithms use the same pairs for fairness. Our tests are performed with 10, 20, 40 and infinite vision ranges and search depths. To test the algorithms, we developed a deliberative off-line prey algorithm (Prey-A*), which is powerful but not very efficient. To prevent the side effects caused by the efficiency difference, the predator and the prey are executed alternately in performance tests. We also assumed that the prey is slower than the predator, and skips 1 move after each 7 moves. With respect to vision ranges and search depths, the averages of path lengths on maze grids are given in Figures 5 and 6, and the averages of path lengths on U-type grids are given in Figures 7 and 8, respectively. In the charts, the horizontal axis is either the vision range or the search depth, and the vertical axis contains the path length
Real-Time Moving Target Search
119
Fig. 6. Average of path length results of maze grids for increasing search depths
Fig. 7. Average of path length results of U-type grids for increasing vision ranges
of MTS-c divided by that of the compared algorithm. The results showed that MTES performs significantly better than MTS-c and MTS-d even with small search depths, and usually offers near optimal solutions that are almost as good as the ones produced by A*. Especially in U-type grids, MTES mostly outperforms A*. When we examined this interesting result in details, we observed that they behave very differently in sparse parts of the grid. MTES prefers performing diagonally shaped manoeuvres for approaching targets located in diagonal directions, whereas A* prefers performing L-shaped manoeuvres in such cases. Since the agents are only permitted to move in horizontal and vertical directions, these two manoeuvre patterns have equal path distances to a static target, but for a moving target, the strategy difference significantly affects the behavior of the prey in U-type grids and sometimes makes A* worse than MTES. We also examined the step execution times of the algorithms running on an AMD Athlon 2500+ computer. In Table 1, the worst case and the average number of moves executed per second are shown. The rows are for the compared algorithms and the columns are for the search depths. We see that MTS-c and MTS-d have low and almost
120
C. Undeger and F. Polat
Fig. 8. Average of path length results of U-type grids for increasing search depths Table 1. Worst case and average number of moves/second for increasing search depths Maze grids Depth 10-sd 20-sd 40-sd MTS-c 1063/2676 1063/2676 1063/2676 MTS-d 937/2412 937/2412 937/2412 MTES 531/1101 212/692 82/413 A* 20/189 20/189 20/168 U-type grids Depth 10-sd 20-sd 40-sd MTS-c 1063/2855 1063/2855 1063/2855 MTS-d 1062/2498 1062/2498 1062/2498 MTES 793/1257 400/747 133/348 A* 8/104 8/104 8/104
constant step execution times whereas the efficiency of MTES is tied to the search depth and obstacle ratio, and hence the appropriate depth should be chosen according to the required efficiency. A* is the worst as expected.
5 Conclusion In this paper, we have examined the problem of pursuing a moving target in grid worlds, introduced a moving target search algorithm, MTES, and presented the comparison results of MTS-c, MTS-d, MTES and A* against a moving target controlled by Prey-A*. With respect to path lengths, the experiments showed that MTES performs significantly ahead of MTS-c and MTS-d, and competes with A*. In terms of step execution times, we observed that MTS-c and MTS- d are the most efficient algorithms, and almost spend constant time in each move. But their solution path lengths are usually unacceptably long. MTES follows MTS, and its efficiency is inversely proportional to the increase in obstacle density. Finally, A* is always the worst.
Real-Time Moving Target Search
121
References 1. Goldenberg, M., Kovarsky, A., Wu, X., Schaeffer, J.: Multiple agents moving target search. In: Int’l. Joint Conf. on Artificial Intelligence, IJCAI, pp. 1536–1538 (2003) 2. Ishida, T., Korf, R.: Moving target search: A real-time search for changing goals. IEEE Trans. Pattern Analysis and Machine Intelligence 17(6), 97–109 (1995) 3. Koenig, S., Likhachev, M.: Real-time adaptive a*. In: 5th Int’l. Joint Conf. on Autonomous Agents and Multiagent Systems, pp. 281–288 (2006) 4. Koenig, S., Likhachev, M., Sun, X.: Speeding up moving-target search*. In: 6th Int’l. Joint Conf. on Autonomous Agents and Multiagent Systems (2007) 5. Korf, R.: Real-time heuristic search. Artificial Intelligence 42(2-3), 189–211 (1990) 6. Russell, S., Norving, P.: Artificial Intelligence: a modern approach. Prentice Hall, Inc., Englewood Cliffs (1995) 7. Stentz, A.: The focussed D* algorithm for real-time replanning. In: Proceedings of the Int’l. Joint Conference on Artificial Intelligence (1995) 8. Undeger, C., Polat, F.: Real-time edge follow: A real-time path search approach. IEEE Transaction on Systems, Man and Cybernetics, Part C 37(5), 860–872 (2007) 9. Undeger, C., Polat, F.: Rttes: Real-time search in dynamic environments. Applied Intelligence 27, 113–129 (2007)
Formalizing Excusableness of Failures in Multi-Agent Systems Eugen Staab and Thomas Engel Faculty of Sciences, Technology and Communication, University of Luxembourg, 6, rue Richard Coudenhove Kalergi, L-1359 Luxembourg Tel.: 00352-466644-5319 {eugen.staab,thomas.engel}@uni.lu
Abstract. To estimate how much an agent can be trusted, its trustworthiness needs to be assessed. Usually, poor performance of an agent leads to a decrease of trust in that agent. This is not always reasonable. If the environment interferes with the performance, the agent is possibly not to blame for the failure. We examine which failures can be called excusable and hence must not be seen as bad performances. Knowledge about these failures makes assessments of trustworthiness more accurate. In order to approach a formal definition of excusableness, we introduce a generic formalism for describing environments of Multi-Agent Systems. This formalism provides a basis for the definition of environmental interference. We identify the remaining criteria for excusableness and give a formal definition for it. Our analysis reveals that environmental interference and a strong commitment of the performing agent do not suffice to make a failure excusable. Keywords: Multi-agent systems, trust, dynamic environments, serviceoriented computing, mobile ad hoc networks.
“Both types of mistake - trusting too well and not well enough - can be costly.” [1]
1
Introduction
In Multi-Agent Systems (MAS) the assessment of agents in terms of trustworthiness is an important topic in research [2, 3, 4, 5, 6]. Many trust-models use experiences made with interaction partners as basis for assessments. But – as Ramchurn et al. point out [2] – only few models account for the context in which the behavior of other agents was assessed. In [7] Falcone and Castelfranchi present a comprehensive study of trust in Multi-Agent Systems. They mention that “when x trusts the internal powers of y, it also trusts his abilities to create positive opportunities for success, to perceive and react to the external problems”. A profound investigation on the causality in the other direction though could not be found in literature, i.e. which experiences should be used to diminish or augment the trust in y. A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 122–133, 2009. c Springer-Verlag Berlin Heidelberg 2009
Formalizing Excusableness of Failures in Multi-Agent Systems
123
An example shows why this is important. Assume the trustworthiness of agent a and agent b is assessed at time t1 and time t2 respectively. Even if the performances of agents a and b are the same, if the environment has changed between time t1 and t2 , agent a may succeed while b fails. The agent who was assessed during the disadvantageous setting is most likely trusted less than the other one. This is an incorrect perception of reality and a sophisticated trust-model should not have this characteristic. We therefore introduce the property of excusableness. If an agent performs well but the environment interferes (such that the agent does not accomplish the task), the failure should not generally be classified as bad performance of the agent; in some cases the failure is excusable. Excusable failures are those which are wrongly used for diminishing the trust in the failing agent. So if the evaluation of experiences is seen as classification problem then excusable failures would belong to the set of false negatives1 . In our work this set is identified by developing a theoretical understanding of excusableness and by providing a definition for it. This definition and all requisite definitions are given from an objective or external viewpoint. In this way, the paper contributes to the conceptual analysis of trust and its assessment. To make the results suitable for the use in formal trust-models, all definitions are presented both informally and formally – by the use of predicate logic. By the definition of excusableness, knowledge about excusable failures can make assessments of trustworthiness more accurate. That is important both for the assessing agent and for the agent being assessed. The paper does not address how this knowledge can be obtained, i.e. how excusableness can be measured in practice. Further we do not investigate moral or legal aspects of the property. For many definitions in this paper we provide examples for the purpose of illustration and to demonstrate the appropriateness of the formalisms. The setting for the examples are Mobile Ad Hoc Networks (MANETs) [8]. A MANET consists of mobile nodes communicating over wireless technologies. The network does not rely on a pre-existing infrastructure as it cannot be predicted how the nodes will move over time. In the next section we introduce a generic formalism for describing environments of Multi-Agent Systems. Using this formalism, environmental interference and riskiness of performances are introduced in Sect. 3. In Sect. 4 the criteria for excusableness are derived and a formal definition is given. We have a look at the work that has been done on related topics in Sect. 5. Conclusions are drawn and an outlook on future work is given in Sect. 6 and 7 respectively.
2
Formalizing the Environment
An environment is described through a set of environmental variables and the relations between them. The set of variables is constructed by taking the power set of all atomic variables where atomic variables are those which are not composed 1
False positives cannot be defined symmetrically and therefore shall be investigated separately.
124
E. Staab and T. Engel
from other variables. Agents are variables, too. Each environmental variable has a certain intensity depending on time. Intensity is a degree of strength (over time) with which the variable influences its environment. The relations between the environmental variables are expressed as their exposure among each other. A variable may be exposed to another one heavily or not at all. The influence one variable has on another one could then be calculated on the basis of intensity of the influencing variable and the exposure of the influenced variable towards the influencing variable. However this will not be an issue of this paper. An environment Ψ can formally be described as a triple Ω, IΩ , EΩ , where: – The set Ω contains all environmental variables ω. The set Ω is constructed as power set of all atomic variables. Each environmental variable is then a set consisting of atomic variables and/or other composed variables. The set of agents (each agent is a variable), denoted by A, is contained in Ω, so A ⊆ Ω. – The set IΩ contains for each variable ω ∈ Ω a function which represents ω’s intensity over time. Every such function ιω : IR → [0, ∞) maps points in time t to the intensity of the environmental variable. The intensity of a variable can range from its minimum 0 (no intensity at all) to ∞. We call a function ιω the intensity-function of variable ω. – The set EΩ contains for each variable ωi ∈ Ω a function which represents the exposure of ωi to all other variables. For a point in time t and an environmental variable ωj such a function εωi ∈ EΩ returns the exposure of ωi to ωj at time t. The function is of the form εω : (Ω ∗ IR) → [0, 1]. A variable can be totally exposed to another one (1) or not at all (0). We call a function εωi the exposure-function of variable ωi . Example 1 (Environment). Let the nodes of a MANET be the set of agents A. As agents are environmental variables each one has an intensity-function which represents the radiation power of an agent’s network card. When an agent a doesn’t send at time t, its intensity-function is zero, i.e. ιa (t) = 0. The exposure between two agents a and b depends on their distance. If their distance changes over time the exposure-functions εa (b, t) and εb (a, t) change (in the same way). So the movement of variables can completely be described by the exposure-functions. Other variables are for instance walls; their intensity-function is constant and depends on the material and the thickness of the wall.
3
Interference and Riskiness
In this section we introduce the concepts of interference and riskiness. Both are important for the definition of excusableness given in Sect. 4. 3.1
Interference with Actions
Interference addresses the impact the environment has on the actions of an agent. More precisely it describes the case in which the environment is the critical cause
Formalizing Excusableness of Failures in Multi-Agent Systems
125
for a failure in an agent’s performance. Critical means that if this cause had not been given, the agent would have succeeded. If there are n environmental n variables ω1 , ..., ωn only together causing a failure, then the variable ω = i=1 ωi interferes with the action. To express interference formally we first need to define two predicates, namely perf orm and success. Definition 1 (Performance). Let A be the set of agents within the set of variables of Ψ and a ∈ A. The predicate perf orm(a, α, t, t , Ψ ) marks the attempt of agent a to perform an action α in an environment Ψ within time-frame [t, t ], i.e. the performance starts at time t and ends at time t . The predicate perf orm does not say how much it was attempted to perform α; so this effort is left unparameterized for the sake of simplicity. It suffices to assume that if this predicate is used more than once in a formula then the effort is the same for each agent for which the predicate is used. Example 2 (Performance). In a MANET an agent attempts to send a packet if it sends the packet with the radiation power of its network card. Definition 2 (Success). Let A be the set of agents within the set of variables of Ψ and a ∈ A. The predicate success(a, α, t, Ψ ) marks the successful accomplishment of an action α through agent a in an environment Ψ at time t. It always holds: success(a, α, t , Ψ ) ⇒ ∃t < t .perf orm(a, α, t, t , Ψ ). Example 3 (Success). An agent in a MANET succeeds to send a packet if it is received without errors. We are now ready to formalize the predicate interf ere(ω, a, α, t, t , Ψ ). Recall that ιω (t) returns the intensity of variable ω at time t. We use the notation IΩ [ιω (t) := y] to denote the substitution of the value of ιω to be y at time t. Definition 3 (Interference of Variables with Actions). An environmental variable ω interferes with the attempt of an agent a to perform an action α within time-frame [t, t ] in an environment Ψ , iff the agent does not succeed but would have succeeded if ω had had a lower intensity. We write interf ere(ω, a, α, t, t , Ψ ) and have interf ere(ω, a, α, t, t , Ψ ) ≡ perf orm(a, α, t, t , Ψ ) ∧ ¬success(a, α, t , Ψ ) ∧∃y < ιω (t), Ψ = Ω, IΩ [ιω (t) := y], EΩ . (perf orm(a, α, t, t , Ψ ) ⇒ success(a, α, t , Ψ )) . The first line of the formula only guarantees that the action is not successfully performed – if it was, there would be no interference. The other two lines enforce the existence of a threshold for the intensity of the interfering variable; if ω had an intensity below this threshold the action would have been successfully accomplished. That excludes the possibility that the agent is not motivated or incapable. Note that the exposure towards the interfering variable needs not to be considered, as the intensity of it can be set to zero, causing the overall influence of the variable to become zero, too (regardless of how much other variables are exposed to it). Analogously a definition could be given using only exposure.
126
E. Staab and T. Engel
Example 4 (Interference). Imagine an agent in a MANET tries to send a packet to another agent. They are separated by a wall which interferes with the transmission. If the wall was not there or maybe only thinner (i.e. the intensityfunction lower), the transmission would have been successful. Another example is the collision of transmissions between two agents using the IEEE 802.11 radio technology and another agent using the Bluetooth radio technology. As they both operate on the 2.4 GHz ISM band an interference is quite possible [9]. If one of the sending agents had not sent (thus the intensity-function would have had value 0) the other transmission would not have been interfered. The definition for interference given above, referred to one specific environmental variable. This can be generalized to apply for an entire environment by quantification over the environmental variables. We use the symbol ∗ as wildcard character: Definition 4 (Interference with Actions). An environment Ψ = Ω, ∗, ∗ interferes with the attempt of an agent a to perform an action α within timeframe [t, t ] in that environment Ψ , iff there is at least one variable ω which interferes with the performance. We write interf ere(a, α, t, t , Ψ ) and have interf ere(a, α, t, t , Ψ ) ≡ ∃ω ∈ Ω.interf ere(ω, a, α, t, t , Ψ ). 3.2
Interference with Tasks
In the previous section a definition for interference of an environment with an atomic action α was given. Now we can proceed to the problem of interference with the accomplishment of a task, e.g. one agent asks another agent to fulfill a task for it. Such a task can consist for instance in delivering products or solving problems. Generally speaking a task can be accomplished by the sequential execution of one or more atomic actions. The interesting thing here is that every agent may have its own way to carry out the task. The environment possibly interferes with only some ways to solve the task but not with others. Note that for a rational agent following the BDI-architecture [10, 11] such a task would be a desire and could be accomplished by the execution of nested plans. A task τ can be fulfilled through the successful execution of a sequence στ of atomic actions. We represent such a sequence στ as a set of triples (t, t , α), 0 t t , associating time-frames with atomic actions, e.g. στ = {(t0 , t0 , α0 ), (t1 , t1 , α1 ), ..., (tn , tn , αn )}. The points of time given in that sequence are not absolute but relative to the point at which the execution of the sequence starts. So if one assumes the above sequence to be chronologically ordered, then t0 = 0. The successful execution of the atomic actions within their associated timeframes leads to the fulfillment of task τ . As we already stated, there can be more than one way to solve a task. The set of all these sequences for one specific τ is written as Sτ . Each agent a has its own set of sequences Sτa for each task τ .
Formalizing Excusableness of Failures in Multi-Agent Systems
127
The definition for interf erence of the environment with a certain sequence στ through an agent a is straightforward (based on definition 4). The only thing to take care of is that the relative dates in the sequence need to be added to the starting point of the execution. Definition 5 (Interference with Sequences). An environment Ψ interferes with the attempt of an agent a to perform a sequence στ in that environment Ψ starting at t0 iff it interferes with at least one atomic action of the sequence. We write interf ere(a, στ , t0 , Ψ ) and have interf ere(a, στ , t0 , Ψ ) ≡ ∃(t, t , α) ∈ στ .interf ere(a, α, t0 + t, t0 + t , Ψ ). Accordingly the predicates perf orm and success of Def. 1 and 2 respectively can be adapted to sequences of actions. We then have perf orm(a, στ , t, t , Ψ ) instead of perf orm(a, α, t, t , Ψ ) and success(a, στ , t, Ψ ) instead of success(a, α, t, Ψ ). Finally it is possible to state in which cases the environment interferes with the attempt of an agent to fulfill a task τ . In order to perform a task τ within a given time-frame, an agent selects a certain sequence στ . This choice is based on observations on the environment, i.e. observations on the intensity- and exposurefunctions of the respective environment Ψ . The choice is represented by the function selecta(τ, t0 , t1 , OΨ ) with domain (T ∗ IR ∗ IR ∗ OΨ ) and the set of according sequences Sτa as codomain; here T denotes the set of all tasks and OΨ the set of all possible sets of observations OΨ on Ψ . Each agent a ∈ A has its own function selecta because every agent may have a different strategy to select a sequence and a different set of sequences Sτa . Definition 6 (Selection). The function selecta (τ, t0 , t1 , OΨ ) returns the selection for a sequence of tasks στ ∈ Sτa an agents makes in order to accomplish a task τ in the time-frame [t0 , t1 ]. Let OΨa,t denote the set of observations on Ψ to which a had access to until time t. Then the environment interferes with the attempt of agent a to fulfill a task τ in an environment Ψ within time-frame [t0 , t1 ], iff the following holds: selecta (τ, t0 , t1 , OΨa,t0 ) = στ ⇒ interf ere(a, στ , t0 , Ψ ). 3.3
Riskiness
Before a sophisticated agent selects a sequence of actions στ in order to fulfill a task τ , it makes a risk assessment for the possible choices. The interest of the agent may be to minimize the risk of a failure (if he wants to appear trustworthy) or to minimize its costs or both. Its risk assessment is based on a set of observations on the environment: Definition 7 (Riskiness). The function risk(a, στ , t0 , OΨa,t0 , Ψ ) returns the probability which an agent a, given observations OΨa,t0 on Ψ , assigns to the event that the sequence of actions στ will interfere with the environment Ψ when executed by agent a at time t0 .
128
E. Staab and T. Engel
Example 5 (Riskiness). Let τ denote the task to securely transfer a packet in a MANET. The probability that this will be accomplished successfully by using a simple substitution cipher for encryption is very low; it saves computation-costs for the sending agent but the risk that it will be compromised is very high. Let’s write στ for this approach. Public key cryptography is computationally more expensive but much more secure [12]. The risk is much lower. We write ςτ for the second approach. For a sufficiently sophisticated agent a we get for a general environment: risk(a, ςτ , t0 , OΨa,t0 , Ψ ) risk(a, στ , t0 , OΨa,t0 , Ψ ).
4
Excusableness
Excusableness identifies cases in which bad performances of agents are wrongly used to decrease the trust in these agents; therefore only failures can be excusable. It can be argued that assessments that account for excusableness are more credulous because certain bad experiences are ignored. In fact, the opposite is true: by ignoring only excusable failures, the difference between agents that are really trustworthy and those which are not is amplified. This has consequences on the social structures of a MAS and, since Castelfranchi et al. [13] identified trust as “relational capital”, agents start to care more about their image as trustee. In the following we analyze the factors that can cause an agent to fail and identify the criteria that make a failure excusable. Consequently the definition of excusableness is given. 4.1
Criteria
An agent can fail in performing a task as a consequence of one or more of the following four factors: – – – –
The ability of the agent, the possibly interfering environment in which the agent performs, the willingness and the commitment of the agent.
Ability If an agent fails to perform a task due to general inability, the trust in that agent should be diminished. If its abilities do not change, the agent will never succeed in the future; other agents should be preferred to it (the agent should be trusted less). When the environment interferes with the performance of one agent, that does not imply that the same environment would interfere with another agent’s actions. One agent might be more robust than another agent and thus should be trusted more; therefore robustness belongs as well to the abilities of an agent. That demands for a definition of excusableness in respect to a reference set of agents. Only if no other agent of the reference set succeeded under the same conditions, a failure might be excusable. This reference set can consist for instance of all agents in a MAS or only the adjacent agents in a network.
Formalizing Excusableness of Failures in Multi-Agent Systems
129
Interference. As listed above, three causes for a failure come from the inside of an agent, namely the ability, the willingness and the commitment of an agent. If these are the only causes for a failure, the reason for failing must be one of reluctance, inability (in respect to the environment) or an insufficient level of commitment. These causes though demand for a decrease of trust in that agent as they can clearly be attributed to the agent itself. The environmental interference is the only external causation and thus must be a precondition for excusableness. As defined in Def. 3, interference implies that the agent is not generally unable to perform the task: the agent would succeed under different circumstances. The ability of the agent itself is thereby already guaranteed if interference is demanded. In a dynamic setting, the trustworthiness of an agent should be independent of the environment in which the agent is assessed. Otherwise, if the environment of the agent changes, the trust in that agent developed earlier ceases to be consistent with the agent’s recent trustworthiness. In a static setting however, the environment of an agent stays always the same. In that case, the agent’s environment should be attributed to the agent itself and wouldn’t make failures excusable. Therefore our definition addresses only dynamic settings. Willingness. A capable agent that is unwilling to perform a task cannot be trusted as it will certainly fail due to reluctance. This directly implies that a failure as a result of unwillingness is not excusable. Willingness is already implicit in the predicate of interference (see Sec. 3.1). With willingness we address the willingness of an agent to perform the task. It does not address how much the performing agent is ready to invest – the next issue: Commitment. An agent can select a way, out of a set of alternative approaches, to accomplish a task τ . Assume the selected approach στ is less costly for the agent, but, to its knowledge, riskier than another approach ςτ ; further assume the agent fails because the approach was risky. Then, of course, the failure is not excusable. As Falcone and Castelfranchi [7] put it: “When x trusts y, x is just assuming that other motivations will prevail over his [y’s] economic interests or other selfish goals”. Therefore, given a measure for riskiness as defined in Def. 7, the definition for excusableness has to require that the performing agent knew no approach that was thought to be less risky and would have succeeded. 4.2
Definition
At this point we are equipped with all the necessary formalisms to give a formal definition of the excusableness-property as derived in the previous section. Note that the selection of a sequence of actions made by the failing agent is based on the observations he had access to until time t0 . To stay comparable, the agents from the reference set need to be equipped with the same set of observations as the failing agent. Recall that εω is the function that describes the exposure of the variable ω towards other variables. Also remember that OΨa,t is the set of observations on Ψ to which a had access to until time t.
130
E. Staab and T. Engel
Definition 8 (Excusableness). The failure of an agent a in performing a task τ in time-frame [t0 , t1 ] and environment Ψ = Ω, IΩ , EΩ is excusable in respect to a group of agents B iff both 1. no other agent in B would succeed in performing τ under the same conditions and 2. for the sequence of actions στ selected by agent a it both holds: (a) στ is interfering with the environment and (b) there is no alternative sequence of actions ςτ with which agent a would have succeeded and which is assessed by a to be less risky than στ . We write excusable(a, B, τ, t0, t1 , Ψ ) with Ψ = Ω, IΩ , EΩ and have excusable(a, B, τ, t0 , t1 , Ψ ) ≡ (1.) ∀b ∈ B, Ψ = Ω, IΩ , EΩ [εb := εa ]. selectb (τ, t0 , t1 , OΨa,t0 ) = στb ⇒ perf orm(b, στb , t0 , t1 , Ψ ) ⇒ ¬success(b, στb , t1 , Ψ ) a,t0 a (2.(a)) ∧ selecta (τ, t0 , t1 , OΨ ) = στ ⇒ interf ere(a, στa , t0 , Ψ ) ∧ ∀ςτ ∈ Sτa .(risk(a, ςτ , t0 , OΨa,t0 , Ψ ) < risk(a, στa , t0 , OΨa,t0 , Ψ) (2.(b)) ⇒ (perf orm(a, ςτ , t0 , t1 , Ψ ) ⇒ ¬success(a, ςτ , t1 , Ψ ))) . Example 6 (Excusableness). Let the four agents a, b, c and d be nodes in a MANET (as shown in Fig. 1). All of them are equipped with the same network cards and send with the same transmitting power. For the purpose of the example they are arranged in a line-topology a-b-c-d, i.e. agent a can only communicate with b, b also with c, d only with c. Now imagine a requests b to deliver a packet to c before time t. Agent b sends the packet repeatedly until time t but c doesn’t receive it as d is sending at the same time: the messages sent by b and d collide and c cannot receive any of them (this is also known as the Hidden Terminal Problem [14]). So b fails in delivering the packet to c. The question is whether this failure should be used by a to diminish its trust in b. That shouldn’t be the case as it could not have been passed off better under these circumstances; this failure should be excusable. Let’s check whether each condition required by Def. 8 is true: 1. No one else would have succeeded: Every packet sent with the same transmitting power would have collided with the packet coming from d (and no other agent could send with a higher transmitting power). 2. (a) The environment did interfere: If the intensity with which the agent d sent would have been 0, then b would have succeeded. (b) There was only one possibility for b to send the packet to c: No less risky alternatives were available to b which would have succeeded.
Fig. 1. The Hidden Terminal Problem in a MANET
Formalizing Excusableness of Failures in Multi-Agent Systems
131
Hence, according to the definition of excusableness it is indeed an excusable failure. In addition, all three conditions are necessary because if any of them was dropped the trust should be diminished and thus the failure should not be classified as excusable: 1. If another agent had succeeded, then a should trust more in that agent than in b concerning the fulfillment of that task. 2. (a) If the environment did not interfere, a failure would imply either b’s inability or its unwillingness: the trust in b should be diminished. (b) If b had sent with a higher transmitting power and would have succeeded in that case, a could expect b to have done so. Agent b’s motivation not to do so would be to save energy – but that is no excuse: agent a should diminish its trust in b.
5
Related Work
The question about excusability is strongly related to the research field of causality (see e.g. Pearl [15]). In the definition of interference (Def. 3) for example, we use counterfactuals: if the environment had been different, the agent would have succeeded – to express this we use two different instances of an environment Ψ of which one is not happening. As the definition of excusableness demands, one precondition is the relativity towards the abilities and commitments of other agents. This draws excusableness out of the field of pure causality. Ramchurn et al. [2] give an overview on the research done in the field of trust in MAS. They point out that many trust models do not take into account in which context assessments of trustworthiness are made. They refer to Molm et al. [16] which state that information captured under certain conditions can become worthless under different conditions. This shows the importance of defining excusableness in dependence of a time-frame [t0 , t1 ] (at another time the failure might not be excusable). In S¸ensoy and Yolum [5], experiences about past interactions are always associated with the environment within which they are made. This enables agents to evaluate their experiences context-sensitively. Similarly Hussain et al. [17] let their trusting peers assimilate recommendations of others based on the context and the time of the according interaction. Both approaches do not account for excusable failures but present frameworks that could be extended accordingly. Falcone and Castelfranchi [7] present a comprehensive study on trust in MultiAgent Systems. They discuss the importance of trust for MAS and analyze the elements forming a mental state of trust. They make a distinction of external and internal factors for trust. With internal factors they mean ability, willingness and commitment. With external they mean the environmental influence. They mention the reciprocal influence between external and internal factors. In a similar direction Boella and van der Torre [18] investigate the motivations of agents when they violate norms. They state that a sophisticated system of trust dynamics would not “decrease the trust if the other agent did its best to fulfill the promise, but failed due to circumstances beyond its control”. The last two
132
E. Staab and T. Engel
papers [7] and [18] agree with our understanding of trust; the authors examine the impact of environmental factors for trusting agents and more general for the phenomenon trust. They do not address the problem of how experiences should be used in the process of developing trust. Furthermore they do not account for the relative aspect of an agent’s ability.
6
Conclusion
In this paper we introduced the notion of excusableness and developed a formal definition for it. A failure shall be called excusable iff it should not be used to diminish the trust in the failing agent. When breaking this proposition down into a set of factors we found out that three conditions must be given such that a failure can be called excusable: 1. No other agent from a reference set of agents would have been able to succeed under the same conditions. 2. The environment has interfered with the performance of the failing agent. 3. No other way to accomplish the task was assessed by the failing agent to be less risky and would have succeeded. It is concluded that environmental interference and a strong commitment of the performing agent do not yet imply excusableness. The relation of the failing agent towards other agents is as well essential in order to make a failure excusable. Further we pointed out that in dynamic settings excusableness has a different meaning than in static settings. In the latter case, the environment of an agent can be interpreted as part of the agent itself. Then environmental interference becomes part of the agent’s inability and doesn’t make a failure excusable. An important question is how knowledge about excusable failures can be obtained in practice. As the definition of excusableness reveals, it is a quite complex property. Thus, in practice it will be hard to gather all the needed information to clearly state whether a failure is excusable or not. This fact promotes either an approach using statistical data analysis and/or a trade-off is made which however still improves the accuracy of trust-assessments.
7
Future Work
Currently we are working on a statistical mechanism that subtracts environmental interference in a reputation system iff it occurred for all assessed agents in the same reference groups. The approach ignores the riskiness-precondition as trade-off; this is acceptable when the reasonable assumption is made that agents do not agree on their risk-strategies over time. The assessing agents collaborate in order to collect as much information as possible. It is planned to verify the improvement within the Agent Reputation and Trust (ART) Testbed Architecture [19].
Formalizing Excusableness of Failures in Multi-Agent Systems
133
References 1. Friedman, B., Kahn Jr., P.H., Howe, D.C.: Trust online. Commun. of the ACM, 34–40 (2000) 2. Ramchurn, S.D., Huynh, T.D., Jennings, N.R.: Trust in multi-agent systems. Knowledge Engineering Review, 1–25 (2004) 3. Huynh, T.D., Jennings, N.R., Shadbolt, N.R.: An integrated trust and reputation model for open Multi-Agent systems. Autonomous Agents and Multi-Agent Systems, 119–154 (2006) 4. Castelfranchi, C., Falcone, R.: Principles of trust for MAS: Cognitive anatomy, social importance, and quantification. In: Proceedings of ICMAS 1998, pp. 72–79 (1998) 5. S ¸ ensoy, M., Yolum, P.: A context-aware approach for service selection using ontologies. In: Proceedings of AAMAS 2006, pp. 931–938 (2006) 6. Teacy, W.T.L., Patel, J., Jennings, N.R., Luck, M.: TRAVOS: Trust and reputation in the context of inaccurate information sources. Autonomous Agents and MultiAgent Systems, 183–198 (2006) 7. Falcone, R., Castelfranchi, C.: Social trust: a cognitive approach. Trust and deception in virtual societies, 55–90 (2001) 8. Haas, Z.J., Deng, J., Liang, B., Papadimitratos, P., Sajama, S.: Wireless ad hoc networks. In: Perkins, C.E. (ed.) Ad Hoc Networking, pp. 221–225. Addison-Wesley, Reading (2001) 9. Vilovic, I., Zovko-Cihlar, B.: Performance analysis of wireless network using bluetooth and IEEE 802.11 devices. In: Proceedings of Elmar 2004, pp. 235–240 (2004) 10. Bratman, M.: Intentions, Plans, and Practical Reason. Harvard University Press (1987) 11. Georgeff, M., Ingrand, F.: Decision-making in an embedded reasoning system. In: Proceedings of IJCAI 1989, pp. 972–978 (1989) 12. Schneier, B.: Applied cryptography, 2nd edn. Protocols, algorithms, and source code. C. John Wiley & Sons, Inc., Chichester (1995) 13. Castelfranchi, C., Falcone, R., Marzo, F.: Being trusted in a social network: Trust as relational capital. In: Stølen, K., Winsborough, W.H., Martinelli, F., Massacci, F. (eds.) iTrust 2006. LNCS, vol. 3986, pp. 19–32. Springer, Heidelberg (2006) 14. Yoo, J., Kim, C.: On the hidden terminal problem in multi-rate ad hoc wireless networks. In: Kim, C. (ed.) ICOIN 2005. LNCS, vol. 3391, pp. 479–488. Springer, Heidelberg (2005) 15. Pearl, J.: Causality: models, reasoning, and inference. Cambridge University Press, Cambridge (2000) 16. Molm, L.D., Takahashi, N., Peterson, G.: Risk and trust in social exchange: An experimental test of a classical proposition. American Journal of Sociology (2000) 17. Hussain, O.K., Chang, E., Hussain, F.K., Dillon, T.S., Soh, B.: Context and time based riskiness assessment for decision making. In: Proceedings of AICT/ICIW 2006, p. 104. IEEE Computer Society, Los Alamitos (2006) 18. Boella, G., van der Torre, L.: Normative multiagent systems and trust dynamics. In: Trusting Agents for Trusting Electronic Societies at AAMAS 2004, pp. 1–17 (2004) 19. Fullam, K., Klos, T., Muller, G., Sabater, J., Topol, Z., Barber, K.S., Rosenschein, J., Vercouter, L.: The agent reputation and trust (ART) testbed architecture. In: Workshop on Trust in Agent Societies at AAMAS 2005, pp. 50–62 (2005)
Design and Implementation of Security Mechanisms for a Hierarchical Community-Based Multi-Agent System Kenichi Takahashi1 , Yoshiki Mitsuyuki2 , Tsunenori Mine2 , Kouichi Sakurai1,2 , and Makoto Amamiya2 1
2
Institute of Systems & Information Technologies/KYUSHU, 2-1-22 Momochihama, Sawara-ku, Fukuoka 814-0001, Japan {takahashi,sakurai}@isit.or.jp Faculty of Information Science and Electrical Engineering, Kyushu University, 744 Motooka, Nishi-ku, Fukuoka 819-0395, Japan {mitsu,mine,amamiya}@al.is.kyushu-u.ac.jp
Abstract. Recently, several community-based systems have been developed; however, almost all such systems have been developed as Webserver-based systems. Thus, server administrator can easily eavesdrop on user communications, since they have to send/receive information through the server. Therefore, we propose multi-agent-based peer-topeer (P2P) system wherein each peer manages his/her information and exchanges it with other peers directly. This, thus, resolves the problems posed by Web-server-based systems; however, we have to consider attacks from malicious third parties. This study designs and implements security protocols/mechanisms for a hierarchical community-based multi-agent system. Furthermore, if we consider a practical use case, we should be able to demonstrate that the proposed system can be implemented by combining it with existing security techniques for more reliable and rapid deployment. Finally, we evaluate the performance of the proposed security system and present an example application.
1
Introduction
The evolution of the Internet has made it difficult to discover target information, further, the nature of the Internet, which allows everyone to easily contribute, has made it difficult to ensure the reliability of information. This imposes the inference of reliability of information on users. Therefore, community-based systems have been developed. In a community-based system, users with common interests and/or objectives organize a community. They can communicate efficiently on the topics within a community. Furthermore, these communities also facilitate the secure exchange of privacy-related information by restricting communications to within a community. Almost community-based systems are developed as Web-server-based systems. These systems enforce security by employing public key infrastructure (PKI), https, access controls mechanisms, etc. However, since users have to send/receive A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 134–145, 2009. c Springer-Verlag Berlin Heidelberg 2009
Design and Implementation of Security Mechanisms
135
information through a server, server administrator can eavesdrop on their communications. Moreover, a vulnerability in the server may threaten the privacy of all users. Therefore, we propose a multi-agent based peer-to-peer (P2P) system wherein each peer manages his/her information and exchanges it with other peers directly. P2P systems can be divided into pure P2P systems such as Gnutella and hybrid P2P systems such as Napster. Hybrid P2P systems employ one or more special peers who provide a special service such as resource discovery and/or indexing of peers, while pure P2P systems do not employ special peers. With regard to a community-based system, a hybrid P2P system is more suitable since special peers organize communities. In the proposed system, the formation of a community depends on a special peer, but each peer in the community can exchange their information directly. This, thus, remedies the problems posed by Web-server-based systems; however, we have to consider attacks from malicious third parties, such as eavesdropping, message alteration, spoofing. In this study, we design and implement security protocols/mechanisms for a hierarchical community-based multi-agent system. Furthermore, if we consider a practical use case, we should be able to demonstrate that the proposed system can be implemented by combining it with existing security techniques for more reliable and rapid deployment.
2
Related Work
In [11], an agent-community-based P2P information retrieval system is proposed; in this system, agent communities manage and look up information efficiently. In [15], a group-based reputation system is proposed, wherein peers evaluate the credibility of a given peer based on his/her local trust information or references from within the group. However, these do not consider security. In [4,13], the security functions for a few multi-agent-based applications have been discussed. In [7], SAgent is proposed for protecting the computations of mobile agent applications in a potentially hostile environment. Also, several researches [10,16] have discussed mobile agent security. These, however, do not take a community structure into consideration. Most security researches in P2P systems have focused on anonymity or trust relationships; however, very few have focused on security countermeasures against attackes such as eavesdropping message alteration, and spoofing [17]. In [1,3] , these security concerns with regard to community-based system have been focused on. [3] allows a peer to change the security policy flexibly based on applications; however, it does not indicate what type of security techniques are required in particular application. [1] focuses on the security on information sharing; however, in this system, a peer has to create a digital signature for each message. Moreover, these studies do not evaluate the performances of the proposed systems. JXTA [14] provides certain functions for the development of secure P2P systems; it requires developers to determine and combine security functions for
136
K. Takahashi et al.
an application. The Globus Project provides the Globus Security Infrastructure (GSI) [6], which is a security framework for grid computing applications. Groove [5] provides spontaneous shared spaces in which a small group of collaborators can securely exchange information. However, these frameworks do not take a hierarchical community structure into considerations.
3
A Hierarchical Community-Based Multi-Agent System
We represent a user as an agent managing his/her information and supporting activities. These agents who have common interests and objectives organize a community. In the proposed model, a community is also defined as one agent; thus, communities can also organize a community. Therefore, our model allows a hierarchical community structure (Fig. 1) [18]. This nature matches that of real societies. For example, each company consists of certain departments with each department comprising certain sections and each section comprises certain staff. As in real societies, a community has a representative, named portal agent. A portal agent is responsible for organizing a community and subsequently providing services for new agent participation, deregistration, and channels to actualize communication from outside the community to within and vice versa. However, the portal agent does not restrict the communications of agents within the community. Thus, agents can communicate with each other freely and directly. This nature protects agents within a community from risks of information leakage caused by a portal agent. 3.1
Levels of Information Security
An agent manages not only non-privacy-related information but also privacyrelated information to support his/her activities. Thus, each agent is responsible for releasing the appropriate information to only the appropriate partner in an appropriate manner. This implies that each agent should protect his/her information by appliying the appropriate cryptographic techniques according Portal E.g., company
Portal
Portal E.g., department
E.g., staff
Agent
Community
Fig. 1. Hierarchical community structure
Design and Implementation of Security Mechanisms
137
to levels of information security. With regard to a system with a hierarchical community structure, we can define four levels of information security: Closed information, which cannot be released to any agent. Partner-limited information, which can be released to only those agents who satisfy a specific condition. Community-limited information, which can be released to only those agents in the same community. Open information, which can be released to all agents. We do not discuss which levels of information security should be applied to a particular information entity, since this depends on the application being considered. Information should be protected in an appropriate manner based on the level of information security. 3.2
Attacks and Corresponding Countermeasures
We have to consider the prevention of attacks from malicious third parties, such as eavesdropping, message alterations and spoofing. Since closed information cannot be released to any agent, we do not need to consider preventive countermeasures for safequarding this type of information against these attacks. However, in the case of the other levels of information security, we have to consider to prevent these attacks. With regard to eavesdropping, no preventive countermeasure are required for open information as it can be released to all agents. However, community -limited and partner-limited information require preventive countermeasures. Community-limited information can be released to only agents in the same community; thus, agents in the same community should be able to access the information but other agents should not. On the other hand, partner-limited information can be released to only those agents who satisfy a particular condition; thus, only an agent who satisfies a specific condition should be able to access the information. Therefore, we introduce the group and P2P keys for encryption/decryption of community-limited and partner-limited information, respectively. In order to protect against message alteration and spoofing, all types of information require preventive countermeasures. We can use a digital signature for this purpose; however, the creation and verification of a digital signatures require high computational power. In this case, we beleive that it would suffice to detect the spoofing when an attack may occur. For example, when we gather price information of a product for the price comparison, influence from information spoofed by a malicious third party is small; but, when we buy the product, the information harms our wallet. Therefore, we propose a mechanism wherein a sender attaches a hash value generated from the content of a message with secret information (e.g. random number) to the message. When the receiver wants to verify whether the message was spoofing or not, he/she requests the secret information (signed by a digital signature) from the sender. On receiving this secret information, if the hash value of the message matches that generated from the secret information, the message can be considered non-spoofed.
138
K. Takahashi et al.
Digital signatures indicate that a particular information entity has been created by someone who possesses a key for its creation; however, it does not indicate who the key belong to. Usually, a certificate authority (CA) keeps a record of which belongs to who by offline communication, ID/password authentication, etc. In the proposed system, a portal agent assumes this role. Therefore, a portal agent specifies requirements for new agent participations. When an agent wishes to join a community, he/she has to provide information that satisfies the specified requirements. Then, the portal agent issues a digital certificate for the agent’s key. The agent can thus prove that the key is recognized by the portal agent by displaying his/her digital certificate. 3.3
Necessary Protocols
We need to design and implement the following protocols for the prevention of attacks from malicious third parties: Participation in a community. It is necessary to prepare for activities of new agents wishing to join a community. Each community has specific requirements for participation, for example, ID/password, invitation from a friend (e.g., mixi [12]), etc. Therefore, a portal agent needs to check whether new agent satisfies the requirements, issue a digital certificate for the new agent’s public key, and provides the group key. Note that the information for satisfying these requirements may also be sensitive. Group key communication. It provides a secure channel for communication among agents from the same community. The sender embeds secret information into a message as the evidence of his/her message. Thus, he/she can prove that the message belongs to him/her. Further, it is also necessary to encrypt and decrypt community-limited information using a group key. Mutual authentication and P2P communication. It provides for the mutual authentication and secure communication between two agents. Mutual authentication is achieved by verifying the digital certificate of each agent. Further, two agents need to share a common key (P2P key) for the encryption and decryption of partner-limited information. After sharing a P2P key, they can exchange partner-limited information securely. Moreover, they can detect whether the group key communication have been spoofed by exchanging secret information. Leaving a community. When an agent leaves a community, his/her digital certificate should be invalidated. Then, his/her digital signature should be recorded in the certificate revocation list (CRL). Group key updation. When community members change, updated group key is necessary for forward and backward security. CRL Reference. Some digital signature may be come invalid before their expiry period. Therefore, an agent may want to refer to the CRL periodically. Although we have designed all of the above protocols, this paper presents only participation in a community, group key communication, and mutual authentication and P2P communication.
Design and Implementation of Security Mechanisms
3.4
139
Security Techniques
We use the following techniques for the design and implementation of the proposed security. Public key cryptosystem. This algorithm utilizes a public-secret key pair. A message encrypted using the public key can only be decrypted using the secret key and vice versa. In this study, the public and secret keys of agent X are depicted as P Kx and SKx , respectively. The encryption and decryption of a message M using a key K are represented as EK (M ) and DK (M ), respectively. Common key cryptosystem. This algorithm utilizes the same key (common key) for message encryption and decryption. Thus, the sender and receiver of a message have to share a common key in advance. Moreover, the algorithm is much faster than the public key cryptosystem. In this study, a common key between agents X and Y is represented as CKxy . Digital signature. A digital signature is created using a secret key (of public key cryptosystem). Since only an agent who knows the secret key can create the digital signature for a message, other agents can confirm that the message is created by the agent who has the secret key. This study represents the digital signature of message M using a secret key SKx as SignSKx (M ). Message authentication code (MAC). MAC is used for the detection of message alteration by using a hash function. The MAC for message M is represented as M AC(M ). Digital certificate. A digital certificate binds a public key with an identity. It is issued by the CA (a portal agent in our system). In this study, the digital certificate created by a portal agent P (who possesses the secret key SKp ) for a public key P Kx is represented as CertSKp (P Kx ). Group key. The group key is shared among the agents belonging to the same community and enables them to encrypt and decrypt a message. In this study, the group key of community C is represented as GKc . Hash chain. A hash chain is a sequence of hash values computed by adapting the hash function n times. In this study, the hash chain with seed S is represented as H 1 (S), H 2 (S), ..., H n (S). Even if H i (S), where i ≤ n, is known to other agents, they cannot compute any H j (S), where j < i.
4
Design of Security Protocols
We designed the protocols for participation in a community, group key communication, and mutual authentication and P2P communication. 4.1
Participation in a Community
With regard to participation in a community, it is necessary to check whether new agent satisfies the requirements for community participation or not, issue a digital certificate for the new agent’s public key, and provides the group key. Therefore, we designed the protocol shown in Fig. 2.
140
K. Takahashi et al. Portal agent P
Agent X PKx
EPKx(CKxp), R, SignSKp(R, CKxp) Verify SignSKp(R, CKxp) Create Q from R
1. When an agent X wishes to participate in a community C, X sends P Kx to the portal agent P of C. 2. P creates a common key CKxp and responds by sending EP Kx (CKxp ), requirements for participation R, and SignSKp (R, CKxp ). 3. X decrypts EP Kx (CKxp ) using SKx and verifies the integrity of CKxp and R using SignSKp (R, CKxp ). Next, X creates a qualification Q for R and sends ECKxp (Q) to P. 4. P decrypts ECKxp (Q) and checks whether Q satisfies R. If it is satisfied, P creates CertSKp (P Kx) and responds by sending CertSKp (P Kx), ECKxp (GKc), and M AC(GKc , CertSKp (P Kx )). 5. X decrypts ECKxp (GKc ) and verifies the possibility of message alteration using M AC(GKc , CertSKp (P Kx)). Also, X creates a hash chain H 1 (Sx), H 2 (Sx), ..., H n (Sx ). Thus, only those agents who satisfy requirements for community participation can obtain the digital certificate and group key. However, in this protocol, X has to know P Kp before the protocol start. If the qualification is one provided from the administrator of C such as ID/password, X can receive P Kp together with ID/password from the administrator. Otherwise, X receives P K? from an agent (who is usually P, but may be a malicious third party) and has to believe P K? is P Kp . If P K? is not P Kp , X may participate in his/her unintended community. However, there is definitely no way to solve this problem, because there is no information to distingush P and a malicious third party. 4.2
Group Key Communication
With regard to group key communication, message encryption using a group key and preventive countermeasures against spoofing are necessary. The proposed system provides two countermeasures against spoofing: One simply involves using a digital signature; however, it requires much amount of computations in order to create a digital signature for each message. The second involves the usage of a hash chain, a sender attaches M AC(M, H n−m (S)) to the m-th message M. Subsequently, the receiver stores the M AC(M, H n−m (S)) and M pair. When the receiver needs to verify whether the message has been spoofed, he/she can do it by obtaining H n−m (S) from the sender (using P2P communication).
Design and Implementation of Security Mechanisms
141
Agent Y
Agent X PKx, CertSKp(PKx) Verify CertSKp(PKy) Create Rx Create CKxy from Rx, RY DCKxy(ECKxy(Ry)) == Ry ? Spoofing check by H(n-k-1)(Sy)
Verify CertSKp(PKx) Create Ry Create CKxy from Rx, Ry DCKxy(ECKxy(Rx)) == Rx ? Spoofing check using H(n-l-1)(Sx)
Fig. 3. Protocol for mutual authentication and P2P communication
4.3
Mutual Authentication and P2P Communication
A protocol for mutual authentication between agents X and Y is shown in Fig. 3. 1. Agent X sends P Kx and CertSKp (P Kx ), which is the digital certificate provided by a portal agent P to agent Y. 2. Y verifies CertSKp (P Kx ) using P Kp . Here, only when Y trusts P, the result of verification evaluates to true. Next, Y generates a random number Ry and responds by sending P Ky , CertSKp (P Ky ), and EP Kx (Ry ). 3. X verifies CertSKp (P Ky ) and generates a random number Rx . Next, the agent X decrypts EP Kx (Ry ) using SKx . Then, X can create a P2P key (common key) CKxy using Rx and Ry . Finally, X sends EP Ky (Rx ), ECKxy (Rx ), and ECKxy (H n−l−1 (Sx )) to Y, where l is the number of messages already sent by X as group key communication. Here, ECKxy (H n−l−1 (Sx )) is optional. It is required only when Y wishes to verify the messages in group key communication have been spoofed. 4. Y decrypts EP Ky (Rx ) using SKy and creates a P2P key CKxy using Rx and Ry . Subsequently, CKxy is surely shared if and only if the decryption of ECKxy (Rx ) yields Rx . Finally, Y responds by sending ECKxy (Ry ) and ECKxy (H n−k−1 (Sy )), where k is the number of messages already sent by Y as group key communication. ECKxy (H n−k−1 (Sy )) is also optional. Moreover, Y can verify whether the messages in group key communication have been spoofed using M AC(M, H n−i−1 (Sx )) generated using H n−i−1 (Sx ), where i > l. 5. X decrypts ECKxy (Ry ). Subsequently, CKxy is surely shared if and only if the decryption of ECKxy (Ry ) yields Ry . X can also verify whether the messages in group key communication have been spoofed. Thus, X and Y share the P2P key CKxy and can exchange partner-limited information securely between them in P2P fashion. 4.4
Considerations of a Hierarchical Community Structure
There are two types of a community structure; upper communities handle important information and lower communities handle important information. For
142
K. Takahashi et al.
Portal P
Exchange PKp, PKq
Portal Q Portal P EGKp(M)
Agent X
ECKpq(M)
Portal Q
EGKq(M)
Agent Y X/Y authenticates Y/X by using PKq/PKp and they share the P2P key CKpq
Fig. 4. Mutual authentication and P2P key sharing between agents beloing to different communities (left), and group key communication between agents belonging to different communities (right)
example, the hierarchical community structure representing a company organization handles important information in executive meetings than one in the meeting in each section. However, the community comprising friends (e.g., mixi) handles important information in the lower communities. Thus, security requirements differ with the type of community structure. A portal agent manages this type of security requirement as a policy. When an upper community handles important information, the portal agent prohibits to forward group key communication messages from the upper community to the lower communities. On the other hand, when lower communities handle important information, the portal agent prohibits to forward group key communication messagess from the lower community to the upper community. Note that if an agent expricitly specifies that message forwarding to an upper/lower community should be allowed, the policy will be neglected. We have to also consider group key and P2P communications between agents belonging to different communities. These communications are kept secure by formulating an agreement between the communities. In the process of agreement, portal agents P and Q of each community exchange each public key P Kp and P Kq , and change their policy to allow message forwarding from the opposite community. Consequently, agents belonging to different communities are able to authenticate each other and share a P2P key. They can, thus, do P2P communications securely (left panel in Fig. 4). Moreover, when lower communities handle important information, group key communication messages between agents belonging to different communities should be protected. Therefore, the portal agents have to share a P2P key. Then, the messages exchanged between the communities are protected by encryption using the P2P key (right panel in Fig. 4). Thus, messages exchanged between agents belonging to different communities are protected.
5
Performance Testing
We implemented the proposed security mechanisms using Java and evaluated their performances.
Design and Implementation of Security Mechanisms
143
Table 1. Results of performance tests (average of 1000 times) Participation Group key communication P2P key sharing with P2P communication mutual authentication 27.93 ms 1.77 ms 35.21 ms 1.52 ms
㪈㪍㪇 㪈㪋㪇
㪫㫀㫄㪼㩷㩿㫄㫊㪀
㪈㪉㪇 㪬㫊㫀㫅㪾㩷㪸㩷㪻㫀㪾㫀㫋㪸㫃 㫊㫀㪾㫅㪸㫋㫌㫉㪼 㪬㫊㫀㫅㪾㩷㪸㩷㪿㪸㫊㪿㩷㪺㪿㪸㫀㫅
㪈㪇㪇 㪏㪇 㪍㪇 㪋㪇 㪉㪇 㪇 㪇
㪌 㪈㪇 㪈㪌 㪉㪇 㪘㩷㫅㫌㫄㪹㪼㫉㩷㫆㪽㩷㪾㫉㫆㫌㫇㩷㫂㪼㫐㩷㪺㫆㫄㫄㫌㫅㫀㪺㪸㫋㫀㫆㫅㫊
Fig. 5. Performances of two anti-spoofing countermeasures (average of 100 times)
In our implementation, we used 1024-bit RSA as the public key cryptosystem, 128-bit AES as the common key cryptosystem, and SHA-256 as the hash function, which are included in Java Cryptography Extention (JCE) [8]. Further, we used the AES key as group key. Our performance tests were conducted on a 100 Base-T local network comprising two PCs running Solaris 10 x86 and Java version 1.5.0 12 with a Pentium4 4GHz CPU and 2GB memory. We set up no requirements for participation, and 128 byte messages were used in group key and P2P communications. The average results of the performance tests are shown in table 1. These results show that 35 agents can participate in the same community, an agent can perform mutual authentication and share P2P key with 28 agents, and an agent can send/receive over 500 messages per second in group key and P2P communication. This implies that in the proposed system, all agents belonging to a community comprising 100 agents can send over 5 multicast messages to the community per second. We also compared the performances of two anti-spoofing countermeasures. The result of this comparison are shown in Fig. 5. When authentication is required after 6 group communications, using a hash chain proves to be more efficient. This is because it reduces the number of digital signature verifications.
6
An Application Example
Nowadays, various business matching services are being actively used and gaining popularity [2,9]. Since these services are developed as Web-server-based systems, all information is stored on the Web site. Thus, companies may not be willing to use sensitive information for business matching. Therefore, a P2P-based system
144
K. Takahashi et al.
is more suitable for business matching services. Therefore, we implemented the proposed security mechanisms using Java and developed a business matching system for small-and-medium-sized companies. In the developed business matching system, the requirements for participation in a community are ID and a password pair check. A company can obtain the ID and password by closing a contract with the administrator of the community. The top community comprises two sub-communities, software and semiconductor community; semiconductor community consists of 12 small communities. An agent participates in the community, exchanges information, and advances business matching semi-automatically. There are two steps in business matching: The first step involves finding candidate business partners from within the community; The second step involves agreeing on terms and contracting the business. The first step requires group key communications for restricting information access to within the community. In the second step, more important information related to the contract is exchanged, therefore, this information should be protected from others by P2P communication. Thus, a company can advance business matching while protecting information.
7
Conclusion
In this study, we have designed and implemented security mechanism for a hierarchical community-based multi-agent system. The proposed system realizes the following: (1) protection of information based on four levels of information security, (2) each community can define unique requirements for participation, and (3) two anti-spoofing methods. Furthermore, we have considered a practical use case, and our security mechanisms can be implemented using Java classes. The results of the performance tests shows that our system allows all agents belonging to a community comprising 100 agents send over 5 multicast messages to the community per second. We believe that this capacity is sufficient for several applications.
Acknowledgments This research was supported by the Strategic Information and Communications R&D Promotion Programme under grant 052310008.
References 1. Berket, K., Essiari, A., Muratas, A.: PKI-Based Security for Peer-to-Peer Information Sharing. In: Proc. of 4th International Conference on Peer-to-Peer Computing, pp. 45–52 (2004) 2. Business Mall, http://www.b-mall.ne.jp/ 3. Detsch, A., Gaspary, L., Barcellos, M., Cavalheiro, G.: Towards a Flexible Security Framework for Peer-to-Peer based Grid Computing. In: Proc. of the 2nd workshop on Middleware for grid computing, pp. 52–56 (2004)
Design and Implementation of Security Mechanisms 4. 5. 6. 7.
8. 9. 10.
11.
12. 13.
14. 15.
16. 17.
18.
145
FIPA MAS Security white paper, http://www.fipa.org/index.html Groove, http://www.groove.net/ GT Security (GSI), http://www.globus.org/toolkit/security/ Gunupudi, V., Tate, S.R.: SAgent: A Security Framework for JADE. In: Proc. of 5th International Joint Conference on Autonomous Agents and Multiagent Systems, pp. 1116–1118 (2006) Java Cryptography Extension (JCE), http://java.sun.com/products/jc Jetro TTPP, https://www3.jetro.go.jp/ttppoas/indexj.html McDonald, T., Yasinsac, A.: Application Security Models for Mobile Agent Systems. In: Proc. of International Workshop on Security and Trust Management, pp. 38–53 (2005) Mine, T., Matsuno, D., Kogo, A., Amamiya, M.: Design and implementation of agent community based peer-to-peer information retrieval method. In: Klusch, M., Ossowski, S., Kashyap, V., Unland, R. (eds.) CIA 2004. LNCS (LNAI), vol. 3191, pp. 31–46. Springer, Heidelberg (2004) mixi, http://mixi.jp/ Poslad, S., Calisti, M., Charlton, P.: Specifying standard security mechanisms in multi-agent systems. In: Falcone, R., Barber, S., Korba, L., Singh, M.P. (eds.) AAMAS 2002. LNCS, vol. 2631, pp. 227–237. Springer, Heidelberg (2003) Project JXTA, http://www.jxta.org/ Tian, H., Zou, S., Wang, W., Cheng, S.: A group based reputation system for P2P networks. In: Yang, L.T., Jin, H., Ma, J., Ungerer, T. (eds.) ATC 2006. LNCS, vol. 4158, pp. 342–351. Springer, Heidelberg (2006) Tsai, J., Ma, L.: Security Modeling of Mobile Agent Systems. J. of Ubiquitous Computing and Intelligence 1, 73–85 (2007) Wallach, D.: A survey of peer-to-peer security issues. In: Okada, M., Pierce, B.C., Scedrov, A., Tokuda, H., Yonezawa, A. (eds.) ISSS 2002. LNCS, vol. 2609, pp. 253–258. Springer, Heidelberg (2003) Zhong, G., Amamiya, S., Takahashi, K., Mine, T., Amamiya, M.: The Design and Implementation of KODAMA System. IEICE Transactions INF. & SYST. E85D(4), 637–646 (2002)
A Need for Biologically Inspired Architectural Description: The Agent Ontogenesis Case Terence L. van Zyl and Elizabeth M. Ehlers Academy of Information Technology, University of Johannesburg, University Rd., Auckland Park, South Africa [email protected], [email protected]
Abstract. Biologically inspired complex adaptive systems (BICAS) have and will continue to move from research laboratories into industry. As the abstractions presented by biologically inspired systems move into industry, systems architects will be required to include the abstractions in their architectural descriptions (ADs) in order to communicate the design to system implementers. The paper argues that in order to correctly present the architectures of BICAS an additional set of biologically inspired views will be required. The paper then describes a set of additional biologically inspired architectural views for use when describing the architecture of BICAS. Finally the paper constructs a set of viewpoints for the biologically inspired views and demonstrates their use in agent ontogenesis. The paper also demonstrates the usage of a number of techniques for describing the architecture of BICAS. Keywords: Architectural Description, Agent Ontogenesis, Biologically Inspired Complex Adaptive Systems, Architectural Views.
A Need for Biologically Inspired Architectural Description
147
The novel contribution of the paper consists of two parts. First an argument is presented in which the short comings of current ADs are highlighted with respect to biologically inspired systems. A set of additional views are presented to facilitate the documenting of biologically inspired complex adaptive agents, a type of biologically inspired complex adaptive system. Secondly an approach to the architectural description of systems that have dynamic architectures that change with time, follow a partially-deterministic path, and are self-adaptive is described. The paper is organised as follows. Section 2 provides background to the topics discussed in the paper. In section 3 a case for biologically inspired ADs is presented. In section 4 an architectural description viewpoint for agent ontogenesis is presented. Finally section 5 looks at related and future work.
2 Background In the background section, the main concepts, terminology, and ideas presented in the paper are defined. The section first explores some concepts relating to the field of biology, then looks at ADs, and finally agent ontogenesis. 2.1 Biology: Genetics, Phylogeny, Ontogeny, and Morphology In biology the field of genetics is primarily concerned with the study of the genetic constitution of an individual, group, or class of living organisms [11], [12]. The term genetic used here refers to those traits of individuals that are affected or determined by their genes [12]. In short a gene is a hereditary unit that determines a particular characteristic or trait in an organism [12], [13]. Characteristics and traits are either morphological (structure) or behavioural. The genes of an individual provide the complete make-up of an organism and are called the individual's genotype [13]. Inspecting the genotype tells what species an organism is and what traits and characteristics are present in the organism. An organism's traits include for example in a fish’s genotype fins and in a dog’s genotype lungs [11], [14]. The environment is a secondary force that has an influence on the resultant traits and characteristics of an organism. The environment's influence is throughout the entire life of an organism and gives rise to the concept of an individual's phenotype. An individual's phenotype is the result of its genotype and environmental factors that together bring about adaptation and change during the development and continued existence of the individual [11], [14]. The genetics driving the genotype and phenotype of an individual is not a random process, but rather is brought about by the historical ancestry of the individual as studied in phylogeny. Phylogeny is the study of, phylogenesis, the evolutionary development of a species or group of organisms through a succession of forms [13]. The succession of forms is brought about by the processes of evolution. The result of phylogenesis is a change in the genetic make-up (genotype) of a species. By changing the genotype phylogenesis allows a species to adapt to changes in its environment. Ontogeny describes the study of the biological process, ontogenesis, that starts with the activation of a single cell and results in a complex mature adult individual
148
T.L. van Zyl and E.M. Ehlers
[15]. In other words ontogeny investigates the “developmental model” of an individual organism from the earliest embryonic stage to maturity [13]. The ontogenesis of an organism can be divided into discreet stages, each of which is characterised by the onset of a particular developmental phenomenon [15]. Ontogenesis is guided by a number of biological processes and structures that express themselves over time. Some of the processes guiding ontogenesis include genetics (discussed above), morphogenesis, and epigenesis (both discussed further below). Epigenesis, relating to epigenetics, is concerned with the expression of genes through interaction with the environment. Epigenesis states that an organism's genotype is not the sole deciding factor in the organism's eventual behaviour or morphology. Epigenetics is concerned with all genetics where the environment plays a role in deciding which genes are expressed and how expression takes place. Examples of epigenetic processes are the central nervous system, immune system, and endocrine system [16]. Many of the learning processes of an individual are shaped by epigenesis. Finally morphology studies those processes concerned with the structure, shape, and form of an individual. Morphogenetics is concerned with how the morphology of an individual comes about. The section has explored some of the biological processes involved in shaping and guiding the evolution and development of biological individuals. In the following section the software engineering practice of architectural description is reviewed. 2.2 Architectural Descriptions (ADs) Software architecture is “the fundamental organisation of a system embodied in its components, their relationships to each other, and to the environment, and the principles guiding the systems design and evolution” [17]. The collection of products that are used to document an architecture are referred to as the architectural description [17]. Fundamental to all software development life cycles is the ability of stakeholders to communicate amongst each other. A system's architectural description provides one such communication tool [17]. An architectural description is composed by combining a collection of products. One of the products used in an architectural description is called a view. A view is a representation of a whole system from the perspective of a related set of concerns [17]. In order to use a view a viewpoint is established [18]. A viewpoint is a pattern or template from which to develop individual views by establishing the purposes and audience for a view and the techniques for its creation and analysis [17]. In summary architecture provides a mechanism, through views and viewpoints, which allows for the production of ADs for software systems. The architectural description guides the software system’s design and evolution [19]. In the following section the concepts outlined in agent ontogenesis are defined. 2.3 Agent Ontogenesis Compositional self-adaptation is dynamic structural self-adaptation of component based software [20]. For the purpose of discussion a software component is briefly described as a “unit of independent deployment and third–party composition” [21]. The actions that constitute compositional self-adaptation include component addition,
A Need for Biologically Inspired Architectural Description
149
component removal, connector addition, connector removal, and component composition. Dynamic here refers to actions that occur at runtime. Component based agent design provides a framework for the design and development of agents from basic building blocks, or components [22], [23]. The authors are researching agent ontogenesis. In agent ontogenesis biologically inspired complex adaptive agents are developed starting from a single component and through compositional self-adaptation and self-organisation arrive at a complex mature developed individual. The actions that constitute compositional self-adaptation and self-organisation are akin to discreet developmental events underlying the developmental model of the agent. It is important to note that the agent is exposed to its environment during its entire developmental process. Being exposed to the environment means that activities such as learning and exploration are present at all stages of the developmental process. Agent ontogenesis describes a method to arrive at a biologically inspired complex adaptive agent. In the next section a case for biologically inspired ADs is presented.
3 The Case for Biologically Inspired Architectural Description The argument presented in the paper shows that current ADs provide inadequate abstractions for the description of BICAS and that additional views are required. The argument first looks at the shortcomings of current ADs. Secondly the field of genetics is described. Thereafter agent ontogenesis is presented as a case that needs additional biologically inspired views. Finally the additional biologically inspired views are presented as viewpoints. 3.1 Shortcomings of Current ADs with Respect to Biologically Inspired Systems The abstractions presented by the various biologically inspired systems provide novel insight into the eventual architecture. The need for biologically inspired views becomes apparent when discussing notion such as supervised learning, genetic programs, or agent ontogenesis. In the case of agent ontogenesis the need for a view that describes the developmental events that result in the adult individual and the continued events that constitute its adaptation until death is apparent. It is natural to describe a neural network in terms of its topology and the architecture of the individual neurons. However if the architecture is communicated to system implementers then the implementation of a neural network from only the above set of views is insufficient. A neural network is not only the result of its static architecture but is also as a result of the training set used in its learning. Both the static structure and the learning mechanism need to be fully described in the architectural description of the neural network. Current ADs are incapable of capturing the additional details required in both of these examples. For a brief expose of current views the reader is referred to [24]. Some might argue that the views presented in popular methodologies such as Krutchen 4+1, reference model for open distributed processes (RM-ODP), and rational unified process (RUP) provide sufficient abstractions to include BICAS [19], [25], [26]. In designing complex systems, architects use abstractions as a mechanism
150
T.L. van Zyl and E.M. Ehlers
to filter unwanted details. The abstractions allow architects to view systems at various levels of complexity and using various idioms. Abstractions are later translated into concrete and well understood patterns as is the case with pipes and filters or service oriented architectures [27]. The use of abstractions become ever more important when designing the architecture of BICAS. BICAS use abstractions such as animats, agents, neurons, and genetic algorithms [2]. Abstractions are essential to the understanding of the goals of the architecture. An example of the abstraction to concrete concept translation is seen in multi-agent systems (MAS). Many MAS implementations, implement agents as objects despite fervent denial by the practitioners that objects and agents are one and the same [28], [29]. In order to take advantage of an agent based system, the abstraction of agent is required. In the section biologically inspired architectural views are explored. 3.2 The Biologically Inspired Views The views presented in the section should not be taken in isolation but should be used in addition to the already established views of other software design methodologies such as kruchten 4+1, RUP, or RM-ODP [19], [25], [26]. The set of views proposed here should be used in situations where the need for abstractions are essential or where the views presented are insufficient to describe the required architecture. 3.2.1 The Phylogenetic View The phylogenetic view is the architectural view relating to, or characteristic of phylogenesis. The phylogenesis of an agent refers to the development of a genotype of the agent through a succession of forms. Typical biologically inspired systems that are described using the phylogenetic view include genetic algorithms and genetic programs. In summary the phylogenetic view is not concerned with the architectural dynamics of an individual but rather with the architectural dynamics as one individual evolves into another individual from generation to generation. 3.2.2 The Ontogenetic View The ontogenetic view describes the ontogenesis of the agent. The view describes a sequential time series of developmental events. Each developmental event triggers a dynamic reconfiguration of the agent's components. The actions that constitute reconfiguration include component addition and subtraction, connection addition and subtraction, and various composition actions. Artificial embryology is an example of a BICAS that requires an ontogenetic view to describe the developmental process. RMODP, Kruchten 4+1, and RUP do not include a view where time series developmental events can be modelled [19], [25], [26]. In summary the ontogenetic view is used to describe how the architecture of an individual changes with respect to time. 3.2.3 The Epigenetic View The epigenetic view describes the epigenesis of a system. The epigenesis of a system involves all those aspects of a system where the type, organisation, behaviour and structure of the components and connectors that constitute the architecture of a system are influenced by environmental feedback. The types of BICAS that would require the epigenetic view include neural networks, and swarm intelligence. In the epigenetic
A Need for Biologically Inspired Architectural Description
151
view those features of a system, relating to self-organisation and self-adaptation are modelled. Both self-organisation and self-adaptation rely on environmental feedback to shape the system's architecture. In summary the epigenetic view describes how the architecture of an individual changes with respect to changes in the environment. 3.2.4 The Morphogenetic View The morphogenetic view is an architectural view used for describing the structure of the components and connectors that make-up an agent. The morphogenetic view describes the static structure of an agent's components and connectors and also how the structure can change over time. All the components that make up an agent refer to the genotype of the agent. Each component can be designed using more common views. In summary the morphogenetic view describes the structure of the individual agent and how that structure can change. 3.2.5 The Environmental View The environment is a first order abstraction in many BICAS [30], [31], [32]. It includes all the entities outside of the agent that the agent is able to receive feedback from. The feedback includes not only software and hardware infrastructure and middleware but also other agents. In the case of reinforcement learning the rewards will be described in the environmental view. Multi agent systems use the environment for communication amongst other functionality [31]. The section has explored a number of additional views not available in common ADs. In the following section the use of some of these views are complemented with viewpoints and examples.
4 Agent Ontogeny, a Viewpoint Example To model a system using a specific view requires the establishment of an associated viewpoint [17]. The associated viewpoint provides the mechanism for constructing and using a view. The section describes some of the biologically inspired views outlined in the paper. In order to model a viewpoint a diagram, some text, or a set of diagrams and or text are required to describe the view. Agent ontogenesis described in section 2.3 is an example of a biologically inspired complex adaptive agent. The architectural description of agent ontogenesis requires the introduction of views not previously required in the description of traditional systems. The need for these additional views stems from the biological abstractions used in designing these agents. Agent ontogeny presents a biologically inspired system that exhibits a dynamic architecture. The architecture presented by agent ontogeny is not only dynamic but is also temporally dependent. In addition to being dynamic and temporally dependent the architecture is also partially-deterministic. The partially-deterministic nature comes as a result of the interaction between the agent and its environment during development. The section uses the example of agent ontogenesis. The viewpoints for the morphogenetic view, ontogenetic view, and epigenetic view are partially declared and an example is given for each view.
152
T.L. van Zyl and E.M. Ehlers
4.1 The Ontogenetic Viewpoint The ontogenetic view is primarily concerned with how the architecture changes over time. Therefore the ontogenetic view requires a time series description of when the architecture will change. In order to describe the ontogenetic view the use of Adaptive-Linear Temporal Logic (A-LTL) is employed as described by [33]. The A-LTL is used to describe the time series adaptation of a telemanufacturing negotiating agent's ontogenesis. A telemanufacturing negotiating agent is an agent that negotiates on behalf of a human for rapid prototyping resources [34]. The example in formula 1 shows an agent starting with a reasoning component and then adding a bidding component. Finally an agent communication language (ACL) component is added as shown in formula 2. The set of adaptation requests are denoted by the time sequence (T0REQ, T1REQ, ... , TnREQ). Where T0REQ denotes the first time interval request up to TnREQ at which stage the agent is fully developed.
Fig. 1. Ontogenetic view of a telemanufacturing negotiating agent's initial development events
Where ^ indicates an “and” operation, ◊ indicates “eventually”, and ⇀ indicates an “adaptation”. REASONINGSPEC is an architectural specification with a single reasoning component and REASONING-BIDDINGSPEC is an architectural specification with reasoning and bidding components. Formula 1 states that the system initially satisfies REASONINGSPEC. After receiving an adaptation request T0REQ the system adapts to satisfy REASONING-BIDDINGSPEC. ADAPT2 = (ADAPT1 ^ ◊T1REQ) ⇀ REASONING-BIDDING-ACLSPEC
(2)
At a later stage the system receives a T1REQ adaptation request. T1REQ causes the system in state ADAPT1 from formula 1 to satisfy REASONING-BIDDING-ACLSPEC. Where REASONING-BIDDING-ACLSPEC indicates an architectural specification that includes reasoning, bidding and ACL components. The graphical representation of the above set of formula can be seen in figure 1. 4.2 The Epigenetic Viewpoint The epigenetic view is primarily concerned with how environmental feedback changes the architecture both parametrically and structurally. The epigenetic viewpoint uses UML state diagrams to describe the adaptation state transitions due to environmental changes. The state diagram representing the agent epigenetic view can be
A Need for Biologically Inspired Architectural Description
153
Fig. 2. Epigentic view of telemanufacturing negotiating agent responding to the environment component Bidder{ provide negotiation dyn; require communication dyn;}
component OptimisedBidder{ provide negotiation dyn; require communication dyn;}
Fig. 4. Optimised Bidder Component component ACL{ provide communication dyn;}
Fig. 6. ACL Component
seen in figure 2. “OPTIMISED-BIDDINGSTATE” and “BIDDINGSTATE” are the states. The environmental events that trigger the transitions between the states are “when (device = slow)” and “when (device = fast)”. 4.3 The Morphogenetic Viewpoint The morphogenetic viewpoint describes the morphogenesis of a system, that is the morphology and how it can change. The morphogenetic viewpoint is described using a Darwin like notation and visualisation [35]. The Darwin notation is informally extended to include an unbind declaration and modes for the purposes of demonstration as described by [36]. The event declaration is also informally included. The event declaration acts in the same way as when however it guards for event states. The choice of Darwin over Dynamic Wright as an architectural description language (ADL) is due to its simplicity [37]. Note that the names of the events triggering reconfigurations correspond to events described in both the epigenetic view and the ontogenetic view. Figure 3 declares a bidding component and figure 4 declares an optimised bidding component, both provide negotiation and requiring communication. In figure 5 a reasoner component is declared that requires negotiation. In figure 6 an ACL component is declared that provides communication. Finally in figure 7 the components are combined to form a TelemanufacturingNegotiating-Agent-Ontogenesis component.
In figure 7, the dotted lines indicate potential binding, the keywords presented in bold have the following meanings: z z z z z z z z
component := component type provide <serviceName> := port for a service provided by a component require <serviceName> := port for a service required by a component inst : := instances of component dyn: dynamic declaration and binding bind .<port>--.<port> := binding between components unbind .<port> := performs the opposite action of bind event <eventName> := waits for specified event then performs operation
For a detailed description of Darwin the reader is referred to [35]. In the following section a comparison to agent oriented software engineering is made.
A Need for Biologically Inspired Architectural Description
155
5 Comparison with Agent Oriented Software Engineering Agent oriented software engineering (AOSE) proposes a methodology for analysing, designing, and building complex software systems. AOSE is an active research area, with many methodologies, that propose engineering software systems using autonomous agents as a point of departure [39], [40], [41]. The field of AOSE has until now focused more on actual architectural solutions, analyses and designs and less on architectural description as put forward in the paper. Further due to this solution driven approach, many AOSE methodologies are biased towards a certain agent design philosophy such as the emphasis on beliefs, desires, and intentions (BDI) highlighted in the paper by Wooldridge [42]. Also apparent in some agent oriented design methodologies is the presentation of abstractions relating mainly to the agent as an entity, agent interactions, and the agent environment [43],[44]. Although the agent, agent interactions, and agent environment are essential for the architectural description of multi-agent systems, the extension to other BICAS are not apparent when notions such as evolution (phylogenesis), ontogenesis, and learning (epigenesis) need to be presented. The architectural views described in the paper are sufficient to describe most views of multi-agent system presented by current AOSE approaches along with any additional view requirements highlighted by the arguments around BICAS presented above. The section has described some of the differences between current AOSE approaches and the ADs put forward by the paper. The authors realise that an approach that augments current AOSE with agent oriented architectural description techniques, such as the use of an ADL, is an open research area that should be explored.
6 Conclusion and Discussion The paper highlights the need for additional formalism and mechanisms for describing the architecture of systems that use BICAS. As these BICAS move from the laboratory into industry the need for mechanisms for describing their architecture will increase. The paper argues that biologically inspired views present the best opportunity for describing BICAS. The benefit of using biologically inspired views is due to the abstractions used by architects being maintained during the architectural process. The paper defines a number of additional views including a phylogenetic view, an ontogenetic view, an epigenetic view, a morphogenetic view, and the environmental view. The paper concludes by using an example of agent ontogenesis for a telemanufacturing negotiating agent and defining a number of viewpoints to describe the architecture of the agent. In the future ADs need to be extended in order to make them sufficiently rich so as to facilitate their use as first class entities in a software systems construction. Model driven architecture (MDA) tries to achieve this richness by providing formalisms and tools surrounding ADs so as to facilitate code generation from models of a system [45]. Future research hopes to use sufficiently rich ADs of a system as the basis for the complete implementation of agent ontogenesis in the way that DNA and genetics provides a complete encoding for living organisms.
156
T.L. van Zyl and E.M. Ehlers
References 1. Gell-Mann, M.: What is Complexity? Complexity, 1 (1995) 2. Cliff, D.: Biologically-Inspired Computing Approaches to Cognitive Systems: A Partial Tour of the Literature. HPL-2003-11 (2003) 3. Luck, M., McBurney, P., Shehory, O., et al.: Agent technology: Computing as interaction (A roadmap for agent based computing) (2005) 4. Andronache, V., Scheutz, M.: Integrating Theory and Practice: The Agent Architecture Framework APOC and its Development Environment ADE, pp. 1014–1021 (2004) 5. Cervenka, R., Greenwood, D., Trencansky, I.: The AML Approach to Modeling Autonomic Systems. ICAS 0, 29 (2006) 6. Faulkner, S., Kolp, M.: Towards an Agent Architectural Description Language for Information Systems 7. Li, J., Mao, X., Shu, Y.: An OO-Based Design Model of Software Agent, pp. 434–440 (2005) 8. Pagliarecci, F., Spalazzi, L., Capuzzi, G.: Formal Definition of an Agent-Object Programming Language. CTS 0, 298–305 (2006) 9. Shan, L., Zhu, H.: Modelling and Specifying Scenarios and Agent Behaviour, p. 32 (2003) 10. Tahara, Y., Ohsuga, A., Honiden, S.: Pigeon: A Specification Language for Mobile Agent Applications. AAMAS 03, 1356–1357 (2004) 11. Hunter, L., Lathrop, R.H.: Guest Editors’ Introduction-Computer Science and Biology: An Unlikely Pair. IEEE Intelligent Systems 17, 8–10 (2002) 12. King, R.C., Stansfield, W.D., Mulligan, P.K.: A dictionary of genetics. Oxford University Press, US (2006) 13. Oxford english disctionary. Oxford University Press, Oxford (2007) 14. Hartl, D., Jones, E.: Genetics: Analysis of Genes and Genomes, 6th edn. Jones & Bartlett (2005) 15. Balon, E. K.: Epigenesis of an Epigeneticist: The Development of some Alternative Concepts on the Early Ontogeny and Evolution of Fishes. Guelph Ichthyology Reviews 1 (1990) 16. Sipper, M., Sanchez, E., Mange, D., et al.: A Phylogenetic, Ontogenetic, and Epigenetic View of Bio-Inspiredhardware Systems. IEEE Transactions on Evolutionary Computation 1, 83–97 (1997) 17. IEEE Recommended Practice for Architectural Description of Software-Intensive Systems. IEEE Std 1471-2000 (2000) 18. Hilliard, R.: Viewpoint Modeling. In: Proceedings of 1st ICSE Workshop on Describing Software Architecture with UML (2001) 19. Kruchten, P.: The 4+1 View Model of Architecture. IEEE Softw. 12, 42–50 (1995) 20. McKinley, P.K., Sadjadi, S.M., Kasten, E.P., et al.: Composing Adaptive Software. Computer 37, 56–64 (2004) 21. Grone, B., Knopfel, A., Tabeling, P.: Component Vs. Component: Why we Need More than One Definition, pp. 550–552 (2005) 22. Rakotonirainy, A., Bond, A.: A Simple Architecture Description Model, p. 278 (1998) 23. Wang, M., Wang, H.: Intelligent Agent Supported Flexible Workflow Monitoring System, pp. 787–791 (2002) 24. Land, R.: A Brief Survey of Software Architecture. Malardalen Real-Time Research Center, Malardalen University, Vasteras, Sweden, Tech. Rep. (2002) 25. Putman, J.R.: Architecting with rm-odp. Prentice Hall PTR (2000)
A Need for Biologically Inspired Architectural Description
157
26. Kruchten, P.: The rational unified process: An introduction. Addison-Wesley Professional (2003) 27. Shaw, M., DeLine, R., Klein, D.V., et al.: Abstractions for Software Architecture and Tools to Support them. IEEE Trans. Softw. Eng. 21, 314–335 (1995) 28. Jennings, N.R.: An Agent-Based Approach for Building Complex Software Systems. Commun. ACM 44, 35–41 (2001) 29. Wooldridge, M.: Intelligent Agents: The Key Concepts, pp. 3–43 (2002) 30. Beer, R.D.: The Dynamics of Active Categorical Perception in an Evolved Model Agent. Adapt. Behav. 11, 209 (2003) 31. Weyns, D., Omicini, A., Odell, J.: Environment as a First Class Abstraction in Multiagent Systems. Auton. Agents Multi-Agent Syst. 14, 5–30 (2007) 32. Okuyama, F.Y., Bordini, R.H., da Costa, R., Carlos, A.: ELMS: An Environment Description Language for Multi-Agent Simulation. In: Environments for Multi-Agent Systems, pp. 91–108 (2005) 33. Zhang, J., Cheng, B.H.C.: Using Temporal Logic to Specify Adaptive Program Semantics. The Journal of Systems & Software 79, 1361–1369 (2006) 34. van Zyl, T.L., Marais, E., Ehlers, E.M.: An implementation of resource negotiating agents in telemanufacturing. In: Sixth International Symposium on Tools and Methods of Competitive Engineering (2006) 35. Magee, J., Kramer, J.: Dynamic Structure in Software Architectures, pp. 3–14 (1996) 36. Kramer, J., Magee, J.: Self-Managed Systems: An Architectural Challenge, pp. 259–268 (2007) 37. Allen, R., Douence, R., Garlan, D.: Specifying and Analyzing Dynamic Software Architectures. Fundamental Approaches to Software Engineering, 21 (1998) 38. Bradbury, J.S., Cordy, J.R., Dingel, J., et al.: A Survey of Self-Management in Dynamic Software Architecture Specifications, pp. 28–33 (2004) 39. Jennings, N.R., Wooldridge, M.: Agent-Oriented Software Engineering. Handbook of Agent Technology. Springer, Heidelberg (2000) 40. Wooldridge, M.J., Jennings, N.R., Kinny, D.: The Gaia methodology for agent-oriented analysis and design. Autonomous Agents and Multi-Agent Systems, 285–312 (2000) 41. Wood, M.F., DeLoach, S.A.: An overview of the multiagent systems engineering methodology. In: Ciancarini, P., Wooldridge, M.J. (eds.) AOSE 2000. LNCS, vol. 1957, pp. 207–221. Springer, Heidelberg (2001) 42. Wooldridge, M., Ciancarini, P.: Agent-Oriented Software Engineering: The State of the Art. In: Ciancarini, P., Wooldridge, M.J. (eds.) AOSE 2000. LNCS, vol. 1957, pp. 1–28. Springer, Heidelberg (2001) 43. Odell, J.J., Van Dyke Parunak, H., Bauer, B.: Representing Agent Interaction Protocols in UML. In: Ciancarini, P., Wooldridge, M.J. (eds.) AOSE 2000. LNCS, vol. 1957, pp. 121– 140. Springer, Heidelberg (2001) 44. Weyns, D., Holvoet, T.: A reference architecture for situated multiagent systems. In: Environments for Multiagent Systems III, 3th International Workshop, E4MAS, Hakodate, Japan (2006) 45. Frankel, D.: Model driven architecture. Wiley, New York (2003)
Multi-Agent Based Web Search with Heterogeneous Semantics Rui Huang1,2 and Zhongzhi Shi1 1
2
Key Laboratory of Intelligent Information Processing Institute of Computing Technology, Chinese Academy of Sciences P.O. Box 2704-28, 100190 Beijing, China Tel.: +86-10-62600508 Graduate School of the Chinese Academy of Sciences, 100049 Beijing, China {huangr,shizz}@ics.ict.ac.cn Abstract. Relevance ranking is key to Web search in determining how results are retrieved and ordered. As keyword-based search does not guarantee relevance in meanings, semantic search has attracted enormous and growing interest to improve the accuracy of relevance ranking. Recently heterogeneous semantic information such as thesauruses, semantic markups and social annotations have been adopted in search respectively for this purpose. However, although to integrate more semantics would logically generate better search results in respect of semantic relevance, such integrated semantic search mechanism is still in absence and to be researched. This paper proposes a multi-agent based semantic search approach to integrate both keywords and heterogeneous semantics. Such integration is achieved through semantic query expansion, meta search of expanded queries in varieties of existing search engines, and aggregation of all search results at the semantic level. With respect to the great volumes of distributed and dynamic Web information, this multi-agent based approach not only guarantees efficiency and reliability of search, but also enables automatic and effective cooperations for semantic integration. Experiments show that the proposed approach can effectively integrate both keywords and heterogeneous semantics for Web search. Keywords: Semantic Search, Multi-Agent System, Relevance Ranking, Semantic Web, Social Annotation.
1
Introduction
Relevance ranking is key to search mechanisms, according to which potentially related documents are retrieved and ordered. As keyword-based search does not guarantee relevance in meanings, search with semantic information, also known as semantic search, has recently attracted enormous and growing research focuses to improve the accuracy of relevance ranking [1,2,3,4].
This work is supported by National Basic Research Programme (No.2007CB311004, No.2003CB317004), 863 National High-Tech Program (No.2006AA01Z128, No. 2007AA01Z132), National Natural Science Foundation of China (No.90604017, No. 60435010, No. 60775035) and Innovation Fund of the Institute of Computing Technology (No.20060250).
A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 158–170, 2009. c Springer-Verlag Berlin Heidelberg 2009
Multi-Agent Based Web Search with Heterogeneous Semantics
159
Heterogeneous semantics have been adopted for search. Latent semantics are induced from statistics of terms in documents, and then used for semantic similarity computing [5]. Thesauruses provide explanations of words and phrases as well as the synonyms and antonyms for search [6,7]. The Semantic Web [8] comprises well-defined semantic markups written in standard languages as RDF1 and OWL2 . Thus it supports semantic inference in search [9,10]. Recent Web 2.0 [11] embraces more social semantics to boost search. For example, the Del.icio.us3 social bookmark services enable search with tags assigned by users as social annotations [12] to their shared web bookmarks. It is expected that to integrate more semantics would logically generate better search results in respect of semantic relevance. As all kinds of semantic information represent some meanings, to understand more of them would generate more proper and intensive semantics for search. It is also demanding to include both semantic and keyword information in search. Keywords and semantics are closely interrelated with each other, and they together represent more exact meanings. Moreover, semantic information is more scarce on the Web compared with keyword information, and the majority of users are inclined to contribute their documents or express their information needs in terms of keywords. However, to the best of our knowledge, no current search mechanism achieves such expected integration of both keywords and heterogeneous semantics. These heterogeneous semantic models differ in their semantic representation standards, relevance ranking criteria, and even suitable search mechanisms. To build a uniform semantic representation and search mechanism is almost impossible, due to the diversified and dynamic semantic information on current open Web. Therefore, a practical approach is to develop an efficient and effective framework to integrate varieties of existing keyword and semantic search engines, and construct a uniform measurement for semantic correlativities and similarities among all keywords and heterogeneous semantics. This paper proposes such a multi-agent based statistical semantic search approach to integrate both keywords and heterogeneous semantics on the Web. A multi-agent architecture is developed for distributed and dynamic integration and cooperation of varieties of existing search engines. Semantic probabilities are defined to calculate semantic relationships among both keywords and heterogeneous semantics. Experiments show that the proposed approach can effectively integrate both keywords and heterogeneous semantics for Web search. Rest of the paper is organized as follows. Section 2 describes the motivation example. Section 3 discusses the related work. Section 4 details the approach with framework overview, fundamental definitions of semantic probabilities, and main algorithms of indexing, knowledge learning, query expansion and results aggregation. Section 5 presents the preliminary experimental results. Conclusions and future works are included in Section 6.
Category (e.g.ODP ) Ontology (e.g.[3]) Tag (e.g.[4])
2
Example Results(not ordered) “Semantic Web - Wikipedia”, semantic web None “W3C Semantic Web”, “Semantic Web Road Map”. S:(adj)semantic. . . “Semantic Web Road Map”, semantic web “semantic analysis” “Latent Semantic Analysis”, S:(n)network,web. . . “Semantic Networks”. categories “Reference: “Semantic Web Road Map”, semantic web . . . Semantic Web”, “Semantic Web: Why RDF “. . . RDF. . . ” is more than XML”. (semantic web) swrc:name, swrc: Tim Berners-Lee ∃isWorkedOn isWorkedOnBy, foaf: , By(Person) :Person, foaf:name Eric Miller. related tags such as “W3C Semantic Web”, semantic web “semantic”,“web”, “FOAF project”, “semanticweb”,“w3c” “Semantic Web Road Map”. Query
Related Semantics
Motivation Example
To illustrate the problem of heterogeneous semantics on the Web, we start with an example webpage “Semantic Web Road Map”4 . Figure 1 illustrates heterogeneous semantics and Table 1 gives semantic search examples for the webpage. In Figure 1, keyword-based indexing (e.g. Google5 ) includes keywords as “semantic” and “web”. LSI(Latent Semantic Indexing) [5] indexes similar terms as “web” and “map”. Look up the WordNet6 thesaurus, we get words and phrases 4 5 6
Multi-Agent Based Web Search with Heterogeneous Semantics
161
as “network ” and “semantic web”. ODP (Open Directory Project 7 ) categorizes the webpage to “Reference:...Semantic Web...” and “...RDF...”. Semantic markups provide more detailed semantics such as “<swrc:name...” specifically describes “semantic web” using the SWRC8 ontology. Del.icio.us users tag the webpage with “semanticweb”, “w3c”, etc. All these heterogeneous semantics represent some meanings of the webpage, which can be used in semantic search. In Table 1, ODP uses categories, [3] uses ontologies and semantic markups, and [4] uses social annotations to boost search. However, we might expect the machine to automatically trace all kinds of related semantics (e.g. “semantics” in Table 1) from simple keyword queries (e.g. semantic web), and integrate all these keywords and heterogeneous semantics search and inference results (e.g. “Example Results” in Table 1). Therefore, a powerful mechanism to integrate both keywords and heterogeneous semantics is needed, in order to make the most of Web information for semantic search. This is the goal of our multi-agent based statistical approach.
3
Related Work
The scientific literature offers various solutions to semantic search, which can be roughly classified into four subcategories according to different semantics used. (1) Latent semantic based methods [5]. (2) Thesaurus based methods, such as for query expansion [6] and similarity computing [7]. (3) Ontology and semantic markup based methods, supporting logical inference of semantic relations [9,10]. Recent efforts focus more on combination of semantic inference and keywordbased search [2,3,13]. (4) Social annotation based methods [4,12,14,15]. Our approach intends to integrate heterogeneous semantics in all these four subcategories. In this sense, it’s different with all current works, though based on them. Two kinds of most related works are [2,3] and [4,12,15]. [2] proposes to tightly integrate semantic inference and text retrieval, which is followed by works as [3] besides ours. However in [2], such integration is achieved only in retrieved results and through feedback mechanisms, while our work automatically expands the query with both related keywords and semantics before search. Our model also differs with [3] which uses centralized mechanisms and fuzzy description logic (DL) to integrate inference and retrieval. We adopt multi-agents and statistical computing to integrate both existing search mechanisms and heterogeneous semantics. [12] enlightens our statistical semantic relationship measurement, yet their work mainly focuses on social relationships of users. Hence we include textual information for analysis rather than pure resource links. Though [15] also explores semantic relationships, it targets at enterprise search and mainly uses annotations as feedbacks, while our approach is designed for the open Web and uses heterogeneous semantics for query expansion and relevance ranking. To integrate varieties of existing search engines, we extend traditional works on multi-agent based meta search mechanisms [16,17,18]. [16] and [18] include 7 8
user model database or user dictionary in order to more exactly understand user intentions (semantics), yet they do not semantically integrate all search results. [17] sheds some light on integration of disparate tourist information and service resources, though the integration is mainly based on ontologies. Our work is among the first to not only expand semantics of queries, but also integrate both keyword and heterogeneous semantic search results. Moreover, we aim at integrating search engines of heterogeneous semantic information, rather than only integrating results of keyword-based search engines. Work on query expansion, detailed in [19], is also related to our work. Characteristics of our spread activation [20] based semantic query expansion is that, we use both keywords and heterogeneous semantics for query expansion. Another relevant research domain is semantic integration, surveyed in [21]. Most current semantic integration is based on certain formalized and structured representations such as data schemas or ontologies [22]. Some pioneer works such as [23] explore emergent semantics for semantic interoperability. To the best of our knowledge, this paper is the first attempt to integrate semantics of different semantic models varied from ontologies to social annotations, though we merely calculate potential semantic correlativities for searching instead of obtaining exact semantic mappings or matchings.
4 4.1
Multi-Agent Based Statistical Semantic Search Overview
Figure 2 illustrates the semantic search system framework based on multi-agent environment MAGE [24]. The main idea is to first calculate semantic probabilities (defined in section 4.2) among all keywords and semantics, second expand the original query with such calculated semantic relationships, then distribute the expanded queries to multiple search engines for search, and finally integrate all search results according to their semantic relevance rankings. In Figure 2, the original document corpus contains some documents of both keywords and heterogeneous semantics, and the semantic knowledge base contains some prior semantic knowledge. The keyword-semantic indexing agent builds the keyword-semantic indexes (see section 4.3). The semantic knowledge learning agent calculates statistical semantic knowledge and enriches the semantic knowledge base (see section 4.4). When a new query is issued or relevance feedbacks are given, the semantic query expansion agent generates a list of expanded queries using keyword-semantic indexes and the semantic knowledge (see section 4.5). This query list is then distributed to varieties of search agents in the system for search and inference. Then the semantic result aggregation agent integrates all keywords and semantics search and inference results, according to their semantic relevance rankings calculated with user-specified modulative aggregation parameters (see section 4.6). If newly found documents accumulate over a threshold number, update the corpus and repeat steps of keyword-semantic indexing and semantic knowledge learning.
Multi-Agent Based Web Search with Heterogeneous Semantics MAGE
Query
Semantic Query Expansion Agent
Expanded Query List
Expanded Query List
Expanded Query List
Expanded Query List
Expanded Query List
... Keyword based Search Agents
... Category based Search Agents
... Thesaurus based Search Agents
... Social Annotation based Search Agents
... Semantic Markup based Search Agents
Search Results
Search Results
Search Results
Search Results
Search Results
163
Relevance Feedbacks
Semantic Knowledge Base
Statistical KeywordSemantic Indexes
Semantic Knowledge Learning Agent
KeywordSemantic Indexing Agent
Document Corpus Semantic Result Aggregation Agent
Statistical Semantic Search
Result Modulative Integration Parameters
Fig. 2. Multi-Agent Based Statistical Semantic Search Framework
4.2
Fundamental Definitions
We mainly consider semantic occurrence, cooccurrence and conditional probabilities. Traditional definitions of probabilities are recasted with respect to the semantic annotation space of both keywords and heterogeneous semantics. By semantic annotation space, we mean the whole set of semantic annotations SA = (sa1 , · · · , sap ). A semantic annotation sa = [T, S] is a pair which defines that the semantic unit (the minimum unit that represents certain complete and clear semantics) list S annotates meanings of the term list T . ∀sj ∈ S semantically occurs in sa, represented as sj Ssa . ∀ti ∈ T semantically occurs in sa, represented as ti Tsa . If sj Ssa ∧ti Tsa , then ti and sj semantically cooccur in sa, represented as < ti , sj >sa . Definition 1 (Semantic Occurrence Probability). ∀t ∈ T , P (t) ∈ [0, 1] denotes the semantic occurrence probability that term t semantically occurs in SA. ∀s ∈ S, P (s) ∈ [0, 1] denotes the semantic occurrence probability that semantic unit s semantically occurs in SA. Definition 2 (Semantic Cooccurrence Probability). ∀ < t, s > where t ∈ T ∧ s ∈ S ∧ SA = [T, S], P (< t, s >) ∈ [0, 1] denotes the probability that term t semantically cooccurs with semantic unit s in SA. Definition 3 (Semantic Conditional Probability). P (t/s) = P () ∈ [0,1] P (s) denotes the semantic conditional probability of term t given semantic unit s, ∈ [0, 1] denotes the semantic conditional probability of seP (s/t) = P () P (t) mantic unit s given term t. Remark 1. Probabilities as P (T ), P (S), P (T /s) P (t/S), P (T /S) can also be defined and inferred from the above definitions and equations, omitted here.
164
R. Huang and Z. Shi
Term-Semantic Inverted Index TermID
Semantic-Term Inverted Index
P ( Term )
SemanticNum
SemanticID P ( Semantic )
TermNum
SemanticID
P ( Term / Semantic )
TermID
P ( Semantic / Term )
SemanticID
P ( Term / Semantic )
TermID
P ( Semantic / Term )
Fig. 3. Keyword-Semantic Indexes Structure
4.3
Statistical Keyword-Semantic Indexing
The keyword-semantic indexes consist of a pair of interrelated indexes: a termsemantic inverted index and a semantic-term inverted index, shown in Figure 3. Term-semantic inverted index is ordered by “TermID” unique to each term, while semantic-term inverted index is ordered by “SemanticID” unique to each semantic unit. P (T erm) is the semantic occurrence probabilities of “Term”. “SemanticNum” stores the length of “SemanticID” links, indicating the total number of related “SemanticID”s. P (T erm/Semantic) is the semantic conditional probability of term T erm given semantic unit Semantic. Similar structure holds for semantic-term inverted index. The two interrelated inverted indexes compose the keyword-semantic indexes of statistical semantic correlativity network. 4.4
Semantic Knowledge Learning
Besides thesaurus and ontologies, the semantic knowledge base also contains learned knowledge about semantic similarity matrixes. These matrixes are learned through Algorithm 1 from the cooccurrence statistics of the document corpus. The idea is similar with LSI [5]. Singular value decomposition (SVD) is applied to cooccurrence probability matrixes for dimension reduction. Then, similarity matrixes are re-constructed from the decomposed matrixes with reduced dimension, representing latent semantic similarities among all terms and semantics. Four such learned matrixes form the semantic similarity knowledge. Algorithm 1. Semantic similarity matrixes learning 1: 2: 3: 4: 5: 6: 7:
Construct original cooccurrence probability matrix M: M[i][j] = P (< T [i], S[j] >) Singular value decomposition: M = UΣV T , where U and V are orthogonal matrixes, and Σ = diag(σ1 , · · · , σr ) is a diagonal matrix Dimension reduction: Σ = diag(σ1 , · · · , σr , 0r +1 , · · · , 0r ) Reconstruct the term-semantic similarity matrix: MT S = UΣ V T , MST = MTTS Compute the term-term similarity matrix: MT T = MT S MST Compute the semantic-semantic similarity matrix: MSS = MST MT S MT S M All four matrixes form the final semantic similarity matrix: M = [ T T ] MST MSS
Multi-Agent Based Web Search with Heterogeneous Semantics
4.5
165
Statistical Semantic Query Expansion
Algorithm 2 for statistical semantic query expansion extends traditional query expansion [19] to use both terms and semantic units. Three steps of semantic computing, semantic inference and query expansion are involved. Algorithm 2. Statistical semantic query expansion Inputs: (a) parsed query Q of (T ype: whether term or semantic unit, ID: term or semantic unit ID, W : weight of the term or semantic unit in query); (b) keyword-semantic indexes IT S , IST ; (c) prior semantic knowledge base KB. Outputs: expanded query Q of (T ype, ID, CP : semantic conditional probability). Constants: (a) θ ∈ [0, 1]: threshold for acceptable semantic conditional probability; (b) κ ∈ [0, 1]: activation decay factor for semantic conditional probability Mid-results: (a) term list T of (ID, CP ); (b) semantic unit list S of (ID, CP ). Step 1. (T, S)← Semantic computing(Q, IT S , IST ); Step 2. (S)← Semantic inference(KB, S); Step 3. (Q )←Query expansion(T, S).
Semantic computing Algorithm 3 finds all related terms and semantic units in the keyword-semantic indexes. From the original query vector, all related keywords and semantic units are traced through iterative calling of subfunction “addList ”. That is, if the decayed (decay each link with κ) semantic conditional probability of a keyword or semantic unit given the query vector exceeds the threshold θ, then the keyword or semantic unit is considered semantically related to the original query and added to the expanded query list. Algorithm 3. (T, S)← Semantic computing(Q, IT S , IST ) 1: for all term +t ∈ Q do 2: addList(t.ID, t.W , 1, θ, κ, T , S, IT S , IST ) 3: end for{terms that should be included in the search results} 4: for all term −t ∈ Q do 5: addList(t.ID, t.W , −1, θ, κ, T , S, IT S , IST ) 6: end for{terms that should be excluded in the search results} 7: for all semantic unit +s ∈ Q do 8: addList(s.ID, s.W , 1, θ, κ, S, T , IST , IT S ) 9: end for{semantic units that should be included in the search results} 10: for all semantic unit −s ∈ Q do 11: addList(s.ID, s.W , −1, θ, κ, S, T , IST , IT S ) 12: end for{semantic units that should be excluded in the search results} SUBFUNCION: addList(int id, float condprob, int include, float threshold, float decayf actor, List L1 , List L2 , Index I1 , Index I2 ) 1: L1 .add(id, include ∗ condprob){When List.add(id, x), if (id, y) already exists, then y ← y + x.} 2: for i = 0 to I1 [id].LinkN um do 3: if (decayf actor ∗ I1 [id].Link[i].CP ) ≥ threshold then 4: addList(I1 [id].Link[i].ID, (decayf actor ∗ I1 [id].Link[i].CP ), include, threshold, (κ ∗ decayf actor), L2 , L1 , I2 , I1 ) 5: end if 6: end for
Semantic inference is processed in Algorithm 4 to query and infer related semantic units using thesauruses and ontologies in the semantic knowledge base. Existing approaches can be applied such as query based on semantic markup languages [25], semantic similarity measure for objects in ontologies [26].
for all semantic unit si ∈ S do for all queried/inferred related semantic unit nsj in KB do if si .CP ∗ MSS [nsj ][si ] ≥ threshold then S.add(nsj .ID, si .CP ∗ MSS [nsj ][si ])) end if {MSS is the semantic-semantic Semantic Similarity Matrix in KB} end for end for
Finally, Algorithm 5 constructs the semantically expanded query with all possibly related terms and semantic units, along with their degree of semantic relativity (normalized to [0, 1]) with the original query. Algorithm 5. (Q )←Query expansion(T, S) 1: 2: 3: 4: 5: 6: 7:
maxCP ← max(max( ti .CP |ti ∈ T ), max( sj .CP |sj ∈ S)) for all term t ∈ T do Q .add(“T erm , t.ID, (t.CP/maxCP )) end for for all semantic unit s ∈ S do Q .add(“Semantic , s.ID, (s.CP/maxCP )) end for
4.6
Search Results Aggregation
Search result aggregation is based on both statistical semantic similarity and inference semantic similarity. Let T = m be the total number of keywords and S = n be the total number of semantic units in the semantic annotations space. The expanded query Q is a 1 ∗ (m + n) matrix q = [qT qS ]. Each result document d ∈ D is a d (m + n) ∗ 1 matrix d = [ T ]. Then statistical semantic similarity of document d dS M M d with query q is calculated as SS(q, d) = qM d = [qT qS ][ T T T S ][ T ]. MST MSS dS Through semantic inference knowledge as well as calculated semantic conditional probabilities, more semantic results can be found through probabilistic inference. Inferred semantic similarity IS(q, d) of document d with query q can be calculated following methods in [27,26]. Therefore semantic similarity of document d with query q is calculated as Sim(q, d) = α∗SS(q, d)+(1−α)∗IS(q, d), where α is a user-specified modulative integration parameter.
5
Experiments
5.1
Datasets
The approach is experimented with heterogeneous semantic datasets on the Web. The WordNet thesaurus of 147,249 unique strings 9 are used not only for 9
http://wordnet.princeton.edu/obtain
Multi-Agent Based Web Search with Heterogeneous Semantics
167
Fig. 4. Semantic search system on MAGE
stemming but also as prior semantic knowledge. ODP data of 730,416 categories and 4,792,967 links obtained May, 2007 10 are included. As for semantic web data, we collect 347 example ontologies on the Web, and use the 10,429,951 RDF triples extracted from Swoogle cache June, 2005 11 . We also crawled a sample of Del.icio.us data during May, 2007, consisting of 459,143 social annotations covering 28,704 different links, 184,136 different users and 54,460 different tags. 5.2
Preliminary Results
Figure 4 is a snapshot of the system running on MAGE. Container-1 holds the statistical semantic search subsystem, while container 2-5 include four different kinds of search engines for both keyword and heterogeneous semantic search. Currently, Google, ODP search, Swoogle and Del.icio.us search are employed. Current experiments included 13,849,511 keywords and 97390 semantic units. Statistics about frequencies of these keywords and semantic units in semantic annotations are:
Such frequencies can not be directly used in probability calculation, as they are too centralized in a small range. Thus we introduce a weighted approach for calculation. W (t Tsa=[T,S] ) ∈ [0, 1] represents weight of t Tsa , W (s Ssa=[T,S] ) ∈ [0, 1] represents weight of s Ssa , and W (< t, s >sa=[T,S] ) ∈ [0, 1] represents weight of < t, s >sa . Then we have: SA
P (t) =
k=1
W (t T sak ) ,
SA
SA
P (< t, s >) = P (t/s) =
P (s/t) =
k=1
SA
P (s) =
k=1
W (s S sak )
SA
W ( sak )
SA
SA W (sa ) k=1 k SA SA S W (ssa ) k=1 k SA SA W (sa ) k k=1 SA SA W (tT ) sak k=1 SA
SA
=
k=1 W ( sak ) SA W (s S sak ) k=1
SA
=
k=1 W ( sak ) SA T k=1 W (t sak )
Different weight functions can be chosen for different strategies to interpret the degree of semantic similarity in social annotations. According to the statistics of keywords and semantics in our experiments, we choose weight function 1 W eight(f req) = lg (f req e ∗ 9 + 1) where f req ∈ [0, 1] is the frequency of the term or semantic unit in the semantic annotation. For the example query “semantic web” in section 2, our system finds an ordered list of related keywords and semantics, including keywords as “web”, “semantic”, “rdf”, “ontology”, “xml” etc, and semantics as “semantic”, “rdf”, “reference”, “category:topic”, “web2.0”, “srwc:isworkedonby” etc. Therefore the search can be processed on all keywords and semantic information.
6
Conclusion
This paper develops a multi-agent based approach for heterogeneous semantic search on the Web. We construct a multi-agent based system to integrate varieties of existing keyword and semantic search engines, and propose a semanticprobability based statistical measure to integrate semantic relationships among both keywords and heterogeneous semantics. Through semantic query expansion, expanded meta search of multiple search engines, and semantic aggregation of all results, our approach uses both keywords and heterogeneous semantics in search. With respect to the great volumes of distributed and dynamic Web information, the proposed approach not only guarantees efficiency and reliability of search, but also enables automatic and effective cooperations for semantic integration. Preliminary experimental results show that varieties of search mechanisms, as well as both keywords and semantics can be effectively integrated for Web search. In the future, we will focus on evaluation and optimization of semantic similarity computing and semantic relevance rank aggregation algorithms.
Multi-Agent Based Web Search with Heterogeneous Semantics
169
References 1. Guha, R., Mccool, R., Miller, E.: Semantic search. In: Proceedings of WWW 2003, pp. 700–709 (2003) 2. Mayfield, J., Finin, T.: Information retrieval on the semantic web: Integrating inference and retrieval. In: SIGIR 2003 Semantic Web Workshop (2003) 3. Zhang, L., Yu, Y., Zhou, J., Lin, C., Yang, Y.: An enhanced model for searching in semantic portals. In: Proceedings of WWW 2005, pp. 453–462 (2005) 4. Bao, S., Wu, X., Fei, B., Xue, G., Su, Z., Yu, Y.: Optimizing web search using social annotations. In: Proceedings of WWW 2007, pp. 501–510 (2007) 5. Furnas, G.W., Deerwester, S., Dumais, S.T., et al.: Information retrieval using a singular value decomposition model of latent semantic structure. In: Proceedings of SIGIR 1988, pp. 465–480 (1988) 6. Voorhees, E.M.: Query expansion using lexical semantic relations. In: Proceedings of SIGIR 1994, pp. 61–69 (1994) 7. Tollari, S., Glotin, H., Maitre, J.L.: Enhancement of textual images classification using segmented visual contents for image search engine. Multimedia Tools and Applications 25(3), 405–417 (2005) 8. Berners-Lee, T., Hendler, J., Lassila, O.: The semantic web. Scientific American 284(5), 34–43 (2001) 9. Cohen, S., Mamou, J., Kanza, Y., Sagiv, Y.: Xsearch: A semantic search engine for XML. In: Proceedings of VLDB 2003, pp. 45–56 (2003) 10. Ding, L., Finin, T., Joshi, A., Peng, Y., Pan, R., Reddivari, P.: Search on the semantic web. IEEE Computer 10(38), 62–69 (2005) 11. O’Reilly, T.: What is web 2.0: Design patterns and business models for the next generation of software. O’Reilly, Sebastopol (2005), http://www.oreilly.com/ 12. Wu, X., Zhang, L., Yu, Y.: Exploring social annotations for the semantic web. In: Proceedings of WWW 2006, pp. 417–426 (2006) 13. Rocha, C., Schwabe, D., de Aragao, M.P.: A hybrid approach for searching in the semantic web. In: Proceedings of WWW 2004, pp. 374–383 (2004) 14. Hotho, A., J¨ aschke, R., Schmitz, C., Stumme, G.: Information retrieval in folksonomies: Search and ranking. In: Sure, Y., Domingue, J. (eds.) ESWC 2006. LNCS, vol. 4011, pp. 411–426. Springer, Heidelberg (2006) 15. Dmitriev, D.A., Eiron, N., Fontoura, M., Shekita, E.: Using annotations in enterprise search. In: Proceedings of WWW 2006, pp. 811–817 (2006) 16. M¨ uller, M.E.: An intelligent multi-agent architecture for information retrieval from the internet (1999) 17. Chiu, D.K.W., Leung, H.-f.: Towards ubiquitous tourist service coordination and integration: a multi-agent and semantic web approach. In: Kishino, F., Kitamura, Y., Kato, H., Nagata, N. (eds.) ICEC 2005. LNCS, vol. 3711, pp. 574–581. Springer, Heidelberg (2005) 18. Koorangi, M., Zamanifar, K.: A distributed agent based web search using a genetic algorithm. International Journal of Computer Science and Network Security 7(1), 65–76 (2007) 19. Khan, M.S., Khor, S.: Enhanced web document retrieval using automatic query expansion. Journal of the American Society for Information Science and Technology 55(1), 29–40 (2004) 20. Crestani, F.: Application of spreading activation techniques in information retrieval. Artificial Intelligence Review 11(6), 453–482 (1997)
170
R. Huang and Z. Shi
21. Noy, N.F.: Semantic integration: a survey of ontology-based approaches. ACM SIGMOD Record 33(4), 65–70 (2004) 22. Gruninger, M., Kopena, J.B.: Semantic integration through invariants. AI Magazine 26(1), 11–20 (2005) 23. Aberer, K., Cudr´eMauroux, P., Hauswirth, M.: The chatty web: emergent semantics through gossiping. In: Proceedings of WWW 2003, pp. 197–206 (2003) 24. Shi, Z., Zhang, H., Cheng, Y., Jiang, Y., Sheng, Q., Zhao, Z.: Mage: An agentoriented programming environment. In: Proceedings of IEEE International Conference on Cognitive Informatics (ICCI 2004), pp. 250–257 (2004) 25. Chu-Carroll, J., Prager, J., Czuba, K., Ferrucci, D., Duboue, P.: Semantic search via xml fragments: A high-precision approach to ir. In: Proceedings of SIGIR 2006, pp. 445–452 (2006) 26. Maguitman, A.G., Menczer, F., Roinestad, H., Vespignani, A.: Algorithmic detection of semantic similarity. In: Proceedings of WWW 2005, pp. 107–116 (2005) 27. Stojanovic, N., Struder, R., Stojanovic, L.: An approach for the ranking of query results in the semantic web. In: Fensel, D., Sycara, K.P., Mylopoulos, J. (eds.) ISWC 2003. LNCS, vol. 2870, pp. 500–516. Springer, Heidelberg (2003)
Reasoning about Norms, Obligations, Time and Agents Jan Broersen1 and Leendert van der Torre2 2
1 University of Utrecht, The Netherlands Computer Science and Communication, University of Luxembourg, Luxembourg
Abstract. Reasoning about norms and time is of central concern to the regulation or control of the behavior of a multiagent system. In this paper we introduce a representation of normative systems that distinguishes between norms and the detached obligations of agents over time, leading to a simple and therefore practical way to reason about norms, obligations, time and agents. We consider the reasoning tasks to determine whether a norm is redundant in a normative system and whether two normative systems are equivalent. In the former case the redundant norm might be removed. In the latter case one norm might be replaced by the other. It is well known that properties concerning iterated or deontic detachment no longer hold when reasoning with multiple agents or with obligations over time. Yet, earlier approaches to reasoning about norms rarely consider the intricacies of time. We show how norms can be used to define the persistence of obligations of agents over time. We illustrate our approach by discussing three ways to relate norms and obligations of agents over time. Also we show how these three ways can be characterized.
1 Introduction and Running Examples Reasoning about norms and time is of central concern to the regulation or control of the behavior of a multiagent system [7, 8]. Norms in multiagent systems come in varying forms, are used for various reasons, and are meant for varying domains. There are norms in law books meant to govern the behavior of communities. Each member of the community is subject to the same norms, and often it is explicitly stated that the norms should not discriminate between members of the society. Contracts constitute another kind of norms. A contract explicitly links agents, for instance a buyer and a seller, to distinct norms. Yet another category is formed by norms describing organizations. Such norms are usually linked to roles used to structure organizations. However, due to the philosophical problem known as Jorgenson’s dilemma [20], which roughly says that a proper logic of norms is impossible because norms do not have truth values, most formal systems reasoning about norms and time [2, 27, 18, 24, 19, 20, 16, 9] are restricted to obligations, prohibitions and permissions only, and do not consider the norms explicitly. These logics might be used, for example, to reason about obligations and permissions following from a legal code without representing the legal code itself. Systems without explicit norms are difficult to use in multiagent systems, but systems that explicitly represent norms of the system usually do not provide a way to reason about them. For example, there is no way to decide whether a norm is redundant in a normative system, or whether two normative systems are equivalent. Ways to reason about the norms [20, 21, 15] do not consider the intricacies of time. A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 171–182, 2009. c Springer-Verlag Berlin Heidelberg 2009
172
J. Broersen and L. van der Torre
We use an example of David Makinson [20] as a running example in this paper. He illustrates the intricacies of temporal reasoning with norms, obligations and agents by discussing the iteration of detachment, in the sense that from the two conditional norms “if p, then obligatory q” and “if q, then obligatory r” together with the fact p, we can derive not only that q is obligatory, but also that r is obligatory. First, Makinson argues that iteration of detachment often appears to be appropriate. He gives the following example, based on instructions to authors preparing manuscripts. Example 1 (Manuscript [20]). Let the set of norms be “if 25x15, then obligatory text12” and “if text12, then obligatory refs10”, where 25x15 is ”The text area is 25 by 15 cm”, text12 is ”The font size for the main text is 12 points”, and refs10 is ”The font size for the list of references is 10 points”. If the facts contain 25x15, then we want to detach not only that it is obligatory that text12, but also that it is obligatory that refs10. Second, he argues that iteration of detachment sometimes appears to be inappropriate by discussing the following example, which he attributes to Sven Ove Hansson. Example 2 (Receipt [20]). Let instances of the norms be “if owejp , then obligatory payjp ” and “if payxy , then obligatory receiptpj ” where owejp is “John owes Peter $1000”, payjp is “John pays Peter $1000”, and receiptpj is “Peter gives John a receipt for $1000”. Intuitively Makinson would say that in the circumstance that John owes Peter $1000, considered alone, Peter has no obligation to write any receipt. That obligation arises only when John fulfils his obligation. Makinson observes that there appear to be two principal sources of difficulty here. One concerns the passage of time, and the other concerns bearers of the obligations. Sven Ove Hansson’s example above involves both of these factors. “We recall that our representation of norms abstracts entirely from the question of time. Evidently, this is a major limitation of scope, and leads to discrepancies with real-life examples, where there is almost always an implicit time element. This may be transitive, as when we say “when b holds then a should eventually hold”, or “. . . should simultaneously hold”. But it may be intransitive, as when we say “when b holds then a should hold within a short time” or “. . . should be treated as a matter of first priority to bring about”. Clearly, iteration of detachment can be legitimate only when the implicit time element is either nil or transitive. Our representation also abstracts from the question of bearer, that is, who (if anyone) is assigned responsibility for carrying out what is required. This too can lead to discrepancies. Iteration of detachment becomes questionable as soon as some promulgations have different bearers from others, or some are impersonal (i.e. without bearer) while others are not. Only when the locus of responsibility is held constant can such an operation take place.” [20] In this paper we show how to reason about norms in these circumstances. In Section 2 we consider obligations that have to be obeyed immediately. In Section 3 obligations are preserved until they are obeyed. To predict the behavior of a normative multiagent system, in Section 4 we investigate the interaction with belief sets of agents. Finally, in Section 5 we suggest how our approach can be extended to reason about abilities, games, norm revision, and other complex phenomena relevant for temporal reasoning about norms and obligations in multi-agent systems.
Reasoning about Norms, Obligations, Time and Agents
173
2 Instantaneous Obligations Our approach to reason about norms, obligations, time and agents involves three steps. Inspired by an approach advocated by Sergot and Craven [26], we assume a branching temporal structure that represents how propositions change over time, and we use an algorithm that, given the input of the branching temporal structure and a set of norms, yields a labeling of the temporal structure with obligations. Moreover, we reason about the norms to determine norm redundancy or equivalence of normative systems. For the temporal structures we use trees, i.e., possibly infinite branching time temporal structures. The reason is that branching time structures are used most often, where the branches represent uncertainty or alternative actions of the agents, and seem to be the most convenient and flexible to model a wide variety of examples. Definition 1 (Temporal structure). Let L be a propositional language built on a set of atoms P . A temporal structure is a tuple T = N, E, |= where N is a set of nodes, E ⊆ N × N is a set of edges obeying the properties of a tree, and |=⊆ N × L is a satisfaction relation for nodes and propositional formulas of L closed under propositional logic. The following example illustrates a possible temporal structure for the manuscript example. Example 3 (Manuscript, continued from Example 1). The temporal structure in Figure 1 must be read as follows. A circle visualizes a node of the temporal structure, and the words within the circle visualize the propositions which are true at that node. An arrow from one node to another one visualizes that the latter node can be reached from the former. In the temporal structure, the root of the tree is visualized on the left hand side, and four distinct nodes can be reached from it. In all these nodes we have that 25x15, which visualizes that the text area is 25 by 15 cm. Then we have or have not text12, the font size for the main text is 12 points or not, and we have or have not refs10, the font size for the list of references is 10 points or not. For all these four nodes we can reach an additional node visualized on the right hand side of the figure, in which we no longer have 25x15. For the normative system we consider only regulative norms like obligations and prohibitions, since they are the most basic and often used kind of norms. Though normative systems may have a complicated structure including, for example, permissive, constitutive and procedural norms [5, 6], they always contain at least such regulative norms. Extensions with, for example, explicit sanctions, as well as other kinds of norms, are left for further research. Following conventions in input/output logic [21, 22], we write a conditional norm “if i, then o is obligatory” as a pair of propositional formulas (i, o). Definition 2 (Normative system). A norm “if i, then obligatory o” is represented by a pair of formulas of L, and written as (i, o). It is also read as the norm “if i, then forbidden ¬o.” A normative system S is a set of norms {(i1 , o1 ), . . . , (in , on )}. Example 4 (Manuscript, continued from Example 3). The normative system contains two norms and is represented by {(25x15, text12), (text12, refs10)}.
174
J. Broersen and L. van der Torre O(text12) O(refs10) 25x15 text12
Fig. 1. Labeling of the temporal structure using the manuscript normative system {(25x15, text12), (text12, refs10)} with instantaneous obligations. The obligation O(refs10) is detached in all 25x15 nodes, also where text12 is false. None of the obligations persists in time.
Norms are used to detach obligations. The detached obligations are a propositional labeling1 of the temporal structure. Definition 3 (Obligation labeling). An obligation labeling is a function O : N → 2L . The following example illustrates how the norms of the manuscript example 1 are used to label our branching time structure. As is always done in the deontic logic literature, we distinguish between so-called factual and deontic detachment [18]. The former is based on a match between the condition of the norm and the facts, and the latter kind is based on a match between the condition and the obligations. As discussed in the introduction, for the manuscript example we want both kinds of detachment. Example 5 (Manuscript, continued from Example 4). The obligation O(text12) is detached in all nodes where 25x15. This is called factual detachment, since the condition of the norm is factually true in the node. The obligation O(refs10) is detached in all 25x15 nodes, also where text12 is false, because text12 is in O(n). This kind of detachment is therefore called deontic detachment. None of the obligations persist in time. The way we label the temporal structure determines the meaning of the norms. Our semantics can thus be considered a kind of operational semantics. We assume that the obligatory formulas at a node are a deductively closed set, representing the ideal alternative for the node. This logical closure is not visualized in Figure 1. Consequently, since in this paper we focus on temporal reasoning with norms, we do not consider contraryto-duty norms stating what should hold in sub-ideal nodes [12, 13] or dilemmas stating various optimal alternatives for a node [28]. For a discussion on the various problems with such norms and approaches such as priorities to deal with these problems, see [14]. 1
Instead we could have introduced an obligation modality with a standard semantics. However this propositional approach avoids some unnecessary complications.
Reasoning about Norms, Obligations, Time and Agents
175
Definition 4 (Instantaneous norm semantics). The instantaneous norm semantics of a normative system S is the unique obligation labeling O : N → 2L such that for each node n, O(n) is the minimal set such that: 1. for all norms (i, o) and all nodes n, if n |= i or i ∈ O(n), then o ∈ O(n). 2. if O(n) |= ϕ then ϕ ∈ O(n), where |= is logical consequence for propositional logic. We now define how to reason about norms, obligations and time. A norm is redundant when it does not affect the labeling of the temporal structure. Definition 5 (Norm redundancy). In normative system S, a norm (i, o) ∈ S is redundant if and only if for all temporal structures, the obligation labeling of S is the same as the obligation labeling of S \ {(i, o)}. The definition of norm redundancy is used to define equivalence of normative systems. Definition 6 (Equivalence normative systems). Two normative systems S1 and S2 are equivalent if and only if each norm of S1 is redundant when added to S2 , and vice versa. Reasoning about normative systems is illustrated in the manuscript example. Example 6 (Manuscript, continued from Example 5). The norm (25x15, refs10) is redundant in normative system {(25x15, text12), (text12, refs10), (25x15, refs10)} due to deontic detachment. Consequently, {(25x15, text12), (text12, refs10), (25x15, refs10)} and {(25x15, text12), (text12, refs10)} are equivalent normative systems. In general, if a norm can be derived using transitivity from the other rules, then it is redundant. Another property for reasoning about norms is reasoning by cases, saying that if the normative system contains (i1 , o) and (i2 , o), then the norm (i1 ∨ i2 , o) is redundant. The following result characterizes all properties which hold for normative systems. For space reasons, we do not detail the proofs of this and following theorems. Theorem 1 (Redundant norms). In a normative system S, a norm (i, o) ∈ S is redundant under the instantaneous semantics when we can derive it from S \ {(i, o)} using replacement of logical equivalents in input and output, together with the following rules: (i1 , o) (i, o1 ∧ o2 ) (i, o1 )(i, o2 ) (i1 , o)(i2 , o) (i, o1 ), (i ∧ o1 , o2 ) SI WO AN D OR CT (i1 ∧ i2 , o) (i, o1 ) (i, o1 ∧ o2 ) (i1 ∨ i2 , o) (i, o2 ) Proof. Let I(n) = {φ | n |= φ} be the set of propositions true at node n, and let S(I(n)) = {o | i ∈ I(n), (i, o) ∈ S}. Then O(n) = ∪{V | S(I(n)) ⊆ V ⊇ S(V ), V complete}. The result follows from the characterization of out4 in input/output logic [21].
Example 7 (Manuscript, continued from Example 5). Assume that text12 is obligatory too for 20x15. {(20x15, text12), (25x15, text12), (text12, refs10), (25x15, refs10)} is equivalent to {(20x15 ∨ 25x15, text12), (text12, refs10)}, because we can derive (20x15 ∨ 25x15, text12) from (20x15, text12) and (25x15, text12) using OR, and vice versa we can derive (20x15, text12) and (25x15, text12) from (20x15 ∨ 25x15, text12) using SI.
176
J. Broersen and L. van der Torre
3 Persistence: Preserving Obligations Until They Are Fulfilled We now turn our attention to Makinson’s second example involving time and agents. The normative system has the same logical structure as the normative system of the manuscript example, but the intended temporal structure contains sequential nodes. Example 8 (Receipt, continued from Example 2). Consider the temporal structure in Figure 2. From the root node visualized on the left side of the figure we can access only one node, in which “John owes Peter $1000”, owejp . From this node, we can reach three other nodes, one in which “John pays Peter $1000”, payjp , another one in which ”Peter gives John a receipt for $1000”, receiptpj , and a third one in which neither happens. The norms are “if owexy , then obligatory payxy ” (owexy , payxy ), and “if payxy , then obligatory receiptyx ” (payxy , receiptyx ). Here x and y are variables ranging over the set of agents, in the sense that each norm is treated as a set of proposition based norms, for each instance of the agent variables. As discussed in the introduction, we do not want to have iterated or deontic detachment, in the sense that Peter is not obliged to give John a receipt until he has given him the money. Moreover, in contrast to the instantaneous semantics, if John pays Peter, then Peter has to give John the receipt directly or at some point in the future. We postpone the definition of deadlines in the present framework [10] for further research. Example 9 (Receipt, continued from Example 8). Consider the desired labeling of the temporal structure in Figure 2. In the circumstance that John owes Peter $1000 but he has not paid him yet, Peter does not have the obligation to write a receipt. That obligation arises only when John fulfils his obligation by paying Peter. Consequently, we have factual detachment without iterated or deontic detachment. Moreover, for each norm, we have that if the condition is true in a node, then the obligation holds for the node itself. Moreover, if the obligation is not fulfilled, then it also holds for all successor nodes. Again, in addition to factual detachment and persistence, we assume that the obligatory formulas at a node are a deductively closed set. This logical closure is not visualized in Figure 2, and it implies that we do not consider contrary-to-duty norms stating what should hold in sub-ideal nodes or dilemmas stating various optimal alternatives for a node. Definition 7 (Persistent norm semantics). The persistent norm semantics of a normative system S is the unique obligation labeling O : N → 2L such that for each node n, O(n) is the minimal set such that: 1. for all norms (i, o), all nodes n1 and all paths (n1 , n2 , . . . , nm ) with m ≥ 1, if n1 |= i and nk |= o for 1 ≤ k ≤ m − 1, then o ∈ O(nm ) 2. if O(n) |= ϕ then ϕ ∈ O(n) Redundance and equivalence for persistent obligations is analogous to reasoning about instantaneous obligations in Definition 5 and 6. We have the following result.
Reasoning about Norms, Obligations, Time and Agents
177
O(receipt pj ) O(receiptpj ) O(payjp ) receiptpj
* payjp O(receipt pj ) O(receiptpj ) HH j H O(receiptpj ) O(receiptpj ) O(payjp ) O(pay jp ) - owejp O(payjp ) pay jp * HH j receipt pj A H O(payjp ) O(payjp ) H A j H A A A U A O(payjp ) O(payjp )
Fig. 2. Labeling of the temporal structure using the receipt normative system S = {(owexy , payxy ), (payxy , receiptyx )} with persistent obligations. The obligation O(receiptpj ) is not detached in all owejp nodes, only when payjp is true. All the obligations persists in time, until they are fulfilled.
Theorem 2 (Redundant norms). In a normative system S, a norm (i, o) ∈ S is redundant under the persistence semantics when we can derive it from S \ {(i, o)} using replacement of logical equivalents in input and output, together with the following rules: (i1 , o) SI (i1 ∧ i2 , o)
(i, o1 ∧ o2 ) WO (i, o1 )
(i1 , o)(i2 , o) OR (i1 ∨ i2 , o)
The theorem shows that the redundancy for persistent obligations is weaker in two respects from redundancy for instantaneous obligations. First, we no longer have redundancy of transitivity because we no longer have deontic detachment. Second, we no longer have the conjunction rule AND, as illustrated by Example 10. Example 10 (Receipt, continued from Example 9). Consider the normative systems S1 = {(owexy , payxy ∧ receiptyx )}, S2 = {(owexy , payxy ), (owexy , receiptyx )} and S3 = {(owexy , payxy ), (owexy , payxy → receiptyx )}. If n |= owejp ∧payjp ∧¬receiptpj , then we have O(n) is the consequence set of payjp ∧ receiptpj . Hence, at node n we cannot discriminate among the three normative systems. However, we have for any successor node of n that using S1 we have payjp ∧ receiptpj ∈ O(m), because the obligation for payjp ∧ receiptpj persists, using S2 we have only receiptpj ∈ O(m), because we have that only the obligation for receiptpj persists, and using S3 , we have only payjp → receiptpj ∈ O(m), because only the obligation for payjp → receiptpj persists. The example also illustrates that the three normative systems S1 , S2 and S3 are not equivalent.
178
J. Broersen and L. van der Torre
4 Beliefs and Subjective Obligations of Agents In normative multi-agent systems we usually consider a subjective obligation Oa p, which can be read as “Agent a is obliged to see to it that p’. Obviously we can replace the objective modality Op in the previous two sections by such subjective obligations, such that, for example, the norm (owexy , payxy ) detaches the obligation Oj (payjp ) in owejp nodes. This does not affect the definitions or results. Moreover, in multi-agent systems we usually make a distinction between subjective knowledge or belief, represented by Ka p or Ba p and objective knowledge or belief, as represented by operators for common knowledge or common belief. The subjective versions of knowledge and belief are relevant for predicting the behavior of a normative multi-agent system, since agents will react only to obligations they know or believe they have [25]. In this section we therefore consider a subjective obligation Oa p read as “agent a believes to be obliged to see it that p”. We replace the objective truth function |= by a subjective belief function B. We assume that all the agents know the norms of the normative system, and therefore we do not relativize the normative system to the set of agents. Definition 8. Let A be a set of agents. An epistemic temporal structure is a tree N, E, B where N is a set of nodes, E ⊆ N × N is a set of edges obeying the properties of a tree, and B : A × N → 2L is a partial function such that B(a, n) contains at least the tautologies and is deductively closed in L. The following example illustrates the temporal structure and a normative system for a situation with subjective beliefs. Example 11 (Promise, continued from Example 2). Consider the temporal structure in Figure 3, which zooms in on the first step of the receipt example. Instead of considering only the situation in which “John owes Peter $1000”, owejp , we consider also the situation in which “John promised Peter $1000”, promisejp . We assume that John is obliged to pay Peter $1000 in either of these two cases. Moreover, we consider three distinct belief sets of John: either he believes that he owes John $1000, or he promised John $1000, or he believes he either owes or promised $1000 to Peter, without remembering which of these two cases hold. The normative system is {(owexy , payxy ), (promisexy , payxy )}. Definition 9. An agent-node labeling is a function O : A × N → 2L . The labeling procedure for subjective obligations is obtained by replacing references to the facts by references to the belief states of agents. Definition 10 (Persistent subjective norm semantics). The persistent subjective norm semantics of a normative system S is the unique obligation labeling O : A × N → 2L such that for each node n, O(a, n) is the minimal set such that: 1. for all norms (i, o), all nodes n1 and all paths (n1 , n2 , . . . , nm ) with m ≥ 1, if i ∈ B(a, n1 ) and o ∈ B(a, nk ) for 1 ≤ k ≤ m − 1, then o ∈ O(a, nm ) 2. if O(a, n) |= ϕ then ϕ ∈ O(a, n)
Reasoning about Norms, Obligations, Time and Agents
179
) Oj (payjp Oj (payjp ) Bj payjp * Bj owejp Oj (payjp ) HH Oj (payjp ) j H Oj (payjp ) Oj (payjp ) B pay * j jp H j H Bj promisejp A H Oj (payjp ) Oj (payjp ) H A j H A A A U A Bj owejp ∨ promisejp
Fig. 3. Labeling of the temporal structure using the ‘promise’ normative system S = {(owexy , payxy ), (promisexy , payxy )} with persistent subjective obligations. The obligation Oj (payjp ) is detached in Bj owejp nodes and in Bj promisejp nodes, but not in Bj (owejp ∨ promisejp ) nodes. All the obligations persist in time, until they are fulfilled.
Example 12 (Promise, continued from Example 11). The labeling obtained by the algorithm is visualized in Figure 3. The obligation Oj (payjp ) is detached in Bj owejp nodes and in Bj promisejp nodes, but not in Bj (owejp ∨ promisejp ) nodes. All the obligations persist in time, until they are fulfilled. The following theorem shows that redundancy for subjective obligations is weaker than redundancy for persistent obligations, because reasoning by cases no longer holds. Theorem 3 (Redundant norms). In a normative system S, a norm (i, o) ∈ S is redundant under the subjective semantics when we can derive it from S \ {(i, o)} using replacement of logical equivalents in input and output, together with the following rules: (i1 , o) SI (i1 ∧ i2 , o)
(i, o1 ∧ o2 ) WO (i, o1 )
Summarizing, if the norms refer to the belief states of agents rather than the objective facts, we lose reasoning by cases. If reasoning by cases is desired for subjective obligations too, then we have to close normative systems explicitly under the OR rule. In other words, we have to add the norm (owexy ∨ promisexy , payxy ) explicitly to the normative system {(owexy , payxy ), (promisexy , payxy )}.
180
J. Broersen and L. van der Torre
5 Further Research: Extensions for Multi-Agent Systems For the instantaneous semantics we can reuse results obtained within the input/output logic framework and extend the normative system with permissive norms [23], constitutive norms like counts-as conditionals [17] and procedural norms [6]. Moreover, we can introduce hierarchies of normative systems with multiple authorities [1]. To deal with contrary-to-duty reasoning and dilemmas we can use constraints [22]. Moreover, we can extend the language of normative systems to refer to obligations which do not refer to the same moment in time, but which hold in the future, by replacing the propositional language L by, for example, LTL or CTL.2 For example, using a modal operator F to refer to any future node, we may model the receipt example by the normative system {(owejp , F payjp ), (payjp , F receiptpj )}.3 Moreover, we can refer to explicit time using for example nominals in our language referring to specific nodes. We can also extend our norms with an explicit deadline “if i, then obligatory o before d,” (i, o < d) [10]. Another issue are dyadic obligations, which we can introduce to model the situation where obligations might be fulfilled before they are detached. In particular, a property of the persistent obligations is that when Peter gives the receipt to John before John has given him the money, maybe because they are in a long standing relationship and Peter trusts John, or maybe because Peter wrongly believed that Peter already transferred the money, then after Peter gives him the money, the obligation to write a receipt is still detached, and persists indefinitely! This is illustrated by the obligation O(receiptpj ) in the third node on the righthand side of Figure 2. We can define an alternative labeling without this property using a labeling with dyadic obligations O(o|c), read as “o is obligatory in context c.” Of particular interest for normative multiagent systems is the extension of our approach with the abilities or powers of the agents. If we label each edge by a (possibly complex) action of the agents, then we can define power of agents using ATL. In this way, we may be able to solve problem of ATL that the power of an agent is a very strong notion, which in reality it holds only under conditions. Finally, we can extend normative systems with a new kind of norms (i, α), where α refers to an action. In that representation, conditional actions can be represented by O(a; b). We label edges instead of nodes. If we label both nodes and edges, then we have both ought-to-be and ought-to-do, and therefore we have to deal with constraints between them, like the green-green-green constraint of Sergot and Craven [26].
2
3
There are many alternatives to the temporal structures we use in our paper, such as infinite trees only, non-branching or linear structures only, or finite state machines instead of trees. The adaptation of our approach to these other temporal structures is left for further research. Our labeling algorithm solves a problem encountered in approaches to reasoning about obligations over time without an explicit normative system [11, 9]. With an explicit normative system, moreover, Example 10 illustrates that our algorithm implies that obligations persist only when they are detached from an explicit norm. Without an explicit normative system, we can only decide among the three normative systems when payjp and receiptpj contain temporal references which can be used to distinguish among them [11, 9].
Reasoning about Norms, Obligations, Time and Agents
181
6 Summary Reasoning about norms and time is of central concern to the regulation of multiagent system behavior [7, 8]. In particular, for the study of social norms emerging in societies enforcing desirable group behavior, the design of legal norms to meet institutional goals in electronic institutions, the design of organizational norms to structure organizations and regulate agents playing a role in the organization, and the study of norms in contracting in electronic commerce. In this paper we show that the distinction between norms and obligations leads to a simple and therefore practical way to reason about norms, obligations of agents and time, and we illustrate our approach by discussing three ways to relate norms and obligations over time. Also we show how these three can be characterized, generalizing the non-temporal input/output logic framework. Our approach to reasoning about norms, obligations, time and agents takes three steps. 1. We assume a branching temporal structure representing how propositions change over time, where the branches represent either uncertainty or alternative actions of the agents. Such a temporal structure can be generated using a formal language, like, for example, in model checkers such as Mocha [3] or action logics such as the causal calculator [4]. 2. We use an algorithm that, given the input of the branching temporal structure and a set of norms, produces as output a labeling of the temporal structure with obligations. The algorithm determines the meaning of the norms, and it may therefore be considered an operational semantics of the norms. We give formal definitions for the possible labelings, enabling us to say when norms are equivalent or redundant. 3. A norm is redundant in a normative system, if removing the norm from the normative system does not change the labeling produced by the algorithm. Two normative systems are equivalent, when adding a norm of one normative system to the other one makes this added norm redundant. In particular, we consider three ways to generate obligations and study their logics. First we consider an instantaneous norm semantics, where the generated obligations have to be obeyed immediately. In this case, iterated detachment holds, as well as reasoning by cases. The second semantics states that obligations are preserved until they are obeyed. In this case, iterated detachment no longer holds, but reasoning by cases does. Finally, we investigate the interaction with belief sets of particular agents at time points, and we show that reasoning by cases no longer holds. We suggest how our approach can be extended to reason about abilities, games, norm revision, and other complex phenomena relevant for temporal reasoning about norms and obligations in multi-agent systems. Moreover, to study the relation between the logic of norms in this paper and existing temporal deontic logics (also known as logics of normative propositions), we can define temporal deontic logics for our temporal structures, and relate the redundancy of norms to logical properties of these temporal structures.
References 1. Alchourr´on, C.E., Makinson, D.: Hierarchies of regulations and their logic. In: Hilpinen, R. (ed.) New Studies in Deontic Logic, pp. 125–148. Reidel (1981) 2. Alchourr´on, C.E., Bulygin, E.: Normative Systems. Springer, Wien (1971)
182
J. Broersen and L. van der Torre
3. Alur, R., Henzinger, T.A., Mang, F.Y.C., Qadeer, S., Rajamani, S.K., Tasiran, S.: Mocha: Modularity in model checking. In: Y. Vardi, M. (ed.) CAV 1998. LNCS, vol. 1427, pp. 521– 525. Springer, Heidelberg (1998) 4. Artikis, A., Sergot, M., Pitt, J.: Specifying electronic societies with the causal calculator. In: Giunchiglia, F., Odell, J.J., Weiss, G. (eds.) AOSE 2002. LNCS, vol. 2585, pp. 1–15. Springer, Heidelberg (2003) 5. Boella, G., van der Torre, L.: An architecture of a normative system. In: Procs. of AAMAS 2006, pp. 229–231. ACM Press, New York (2006) 6. Boella, G., van der Torre, L.: Substantive and procedural norms in normative multiagent systems. Journal of Applied Logic (2008) 7. Boella, G., van der Torre, L., Verhagen, H. (eds.): Computational and Mathematical Organization Theory. Selected papers from NORMAS 2005 (2006) 8. Boella, G., van der Torre, L., Verhagen, H. (eds.): Normative Mulit-agent systems, Procs. of NORMAS 2007, Dagstuhl Seminar proceedings 07122 (2007) 9. Broersen, J., Brunel, J.: What I fail to do today, I have to do tomorrow: a logical study of the propagation of obligations. In: Proceedings CLIMA VIII. LNCS. Springer, Heidelberg (2007) 10. Broersen, J., Dignum, F., Dignum, V., Meyer, J.-J.: Designing a deontic logic of deadlines. In: Lomuscio, A., Nute, D. (eds.) DEON 2004. LNCS, vol. 3065, pp. 43–56. Springer, Heidelberg (2004) 11. Brunel, J., Bodeveix, J.-P., Filali, M.: A state/Event temporal deontic logic. In: Goble, L., Meyer, J.-J.C. (eds.) DEON 2006. LNCS, vol. 4048, pp. 53–68. Springer, Heidelberg (2006) 12. Chisholm, R.: Contrary-to-duty imperatives and deontic logic. Analyse 24, 33–36 (1963) 13. Forrester, J.: Gentle murder, or the adverbial samaritan. The Journal of Philosophy 81, 193– 197 (1984) 14. Hansen, J., Pigozzi, G., van der Torre, L.: Ten philosophical problems in deontic logic. In: Proceedings of Normative Multi-agent Systems, NORMAS 2007 (2007) 15. Hansen, J.: Sets, sentences, and some logics about imperatives. Fundam. Inform. 48(2-3), 205–226 (2001) 16. Horty, J.: Agency and Deontic Logic. Oxford University Press, Oxford (2001) 17. Jones, A., Sergot, M.: A formal characterisation of institutionalised power. Logic Journal of IGPL (1996) 18. Loewer, Belzer: Dyadic deontic detachment. Synthese 54, 295–318 (1983) 19. Makinson, D.: Five faces of minimality. Studia Logica 52, 339–379 (1993) 20. Makinson, D.: On a fundamental problem of deontic logic. In: McNamara, P., Prakken, H. (eds.) Norms, Logics and Information Systems. New Studies on Deontic Logic and Computer Science, pp. 29–54. IOS Press, Amsterdam (1999) 21. Makinson, D., van der Torre, L.: Input-output logics. Journal of Philosophical Logic 29(4), 383–408 (2000) 22. Makinson, D., van der Torre, L.: Constraints for input-output logics. Journal of Philosophical Logic 30(2), 155–185 (2001) 23. Makinson, D., van der Torre, L.: Permissions from an input-output perspective. Journal of Philosophical Logic 32(4), 391–416 (2003) 24. Meyer, J.J.C.: A different approach to deontic logic: Deontic logic viewed as a variant of dynamic logic. Notre Dame Journal of Formal Logic 29(1), 109–136 (1988) 25. Pacuit, E., Parikh, R., Cogan, E.: The logic of knowledge based obligation. Knowledge, Rationality and Action 149(2), 311–341 (2006) 26. Sergot, M.J., Craven, R.: The deontic component of action language nC+. In: Goble, L., Meyer, J.-J.C. (eds.) DEON 2006. LNCS, vol. 4048, pp. 222–237. Springer, Heidelberg (2006) 27. van Eck, J.: A system of temporally relative modal and deontic predicate logic and its philosophical applications. logique et analyse 25, 339–381 (1982) 28. van Fraassen, B.: Values and the heart’s command. The Journal of Philosophy (1973)
An Agent Modeling Method Based on Scenario Rehearsal for Multiagent Simulation Shohei Yamane and Toru Ishida Department of Social Informatics, Kyoto University Yoshida-Honmachi, Sakyo-ku, Kyoto, 606-8501, Japan [email protected], [email protected]
Abstract. Multiagent Systems are potential computational systems for various practical applications, tools, and so on. Multiagent simulation is one of the remarkable application to evaluate several kinds of phenomena. In order to design an agent for multiagent simulation, it is important to reflect user’s opinion. However, if a user is not computer professional or does not have technical knowledge of agent logics and programming language, it is hard for him/her to implement his/her own opinion. Participatory design is a promising approach to incorporate user’s opinion in the agent design and modification process. In this paper, we propose rehearsal oriented testing for implementation of participatory design. By the rehearsal oriented testing, it becomes possible to carry out anytime modification of agent’s scenario, which describe its behavior during simulation. For rehearsal oriented testing, we set operators for modifying scenarios, which is described using finite state machine model. We also design interaction protocol between a user and an agent to smoothly get information through the user-agent dialog for modifying operators. Under this protocol, an agent informs a user about what kind of information is required.
1
Introduction
Multiagent Systems are potential computational systems for various practical applications, tools, and so on. Multiagent simulation is one of the remarkable application to evaluate several kinds of phenomena. In order to design an agent for multiagent simulation, it is important to reflect user’s opinion. For agents which provide services for humans, the quality of services will be improved by taking account of users’ preference. For agents which simulate users, the agents’ behavior can be similar to them. Participatory design is a promising approach to incorporate user’s opinion in design process. There are some researches which try to apply participatory design to multiagent systems. Participatory simulations are examples of such researches[1][2]. In a participatory simulation, users participate in the simulation as avatars in the simulation. Log data of the simulation and interviews with participants using a videotaped simulation are used to construct agent models[3]. A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 183–194, 2009. c Springer-Verlag Berlin Heidelberg 2009
184
S. Yamane and T. Ishida
Applying users’ opinion is important not only in the designing process but also in the maintenance process. It is ineffective to repeat the process of agent design whenever the modification or extension of agent models are required. If users can modify/extend agents’ behavior by themselves, it is easy to effectively implement their own opinion on agents. However, if a user is not computer professional or does not have technical knowledge of agent logics and programming language, it is hard for him/her to implement his/her own opinion. This research proposes rehearsal oriented testing so that users who are not computer professionals can modify/extend agents’ behavior. In this research, we propose rehearsal oriented testing to help users, who are not computer professionals or do not have technical knowledge of agent logics and programming language, to modify and extend agents’ behavior. In this paper, we use metaphor of rehearsal for drama in order to construct rehearsal oriented testing methodology. A user modifies/extends agents’ behavior as a director directs actors. In this paper, we addresses the following two issues to realize rehearsal oriented testing. Setting of modification operators. We should reveal targets of modification and then we set modification operators required to modify them. Design of an interaction protocol. In order to modify/extend an agent’s behavior through dialog between a user and an agent, we design an interaction protocol based on defined modification operators. This paper is organized as follows. Section 2 explains what users do in rehearsal oriented testing. Section 3 explains a interaction protocol between a user and an agent for rehearsal oriented testing. Section 4 explains possible applications of rehearsal oriented testing. Section 5 concludes this paper with our future work.
2
Rehearsal Oriented Testing
Our main motivation for introducing rehearsal oriented testing is modification of agents’ behavior which designed based on participatory simulation approach. Participatory simulations incorporate avatars which are characters operated by humans as in computer games in multiagent simulations. Agents which simulate humans are designed on the basis of log data of avatars and interviews to the humans. In this case, users of rehearsal oriented testing are human subjects of participatory simulations. The user monitors the agent’s behavior which simulates him/her and modifies the agent’s behavior if it is different from the user’s expectation. In rehearsal oriented testing, each user plays a role as a director and an agent play role as an actor. Monitoring agents, a user modifies their behavior like a director supervises actors. Even if a user is not a computer professional, he/she can modify agents’ behavior by rehearsal oriented testing. Figure 1 shows an overview of an interface for rehearsal oriented testing. To perform rehearsal oriented testing, a simulation environment should be
An Agent Modeling Method Based on Scenario Rehearsal
185
Fig. 1. Interface for rehearsal oriented testing
visualized in order for a user to monitor agents’ behavior. In Figure 1, we used FreeWalk/Q1 [4] to visualize the simulation. A user can check whether there are problems in an agent’s behavior through monitoring of the visualized simulation. When a user finds a problem, he/she interrupts the agent’s behavior then starts interaction in order to modify the ongoing behavior. That is to say, a user tries to conduct modification. In Figure 1, an agent opens a window which displays the agent’s remark by text and information available for the user. In this example, the agent is asking its user to specify what it should do to fix the problem next and showing its plan described in state transition diagram. We explain more detail about interaction between a user and an agent in the rest of this section. The user gives direction to the agent based on given information through the window by typing a text. For example, assume a case that a user tells his/her agent to turn around because it is getting stuck at the corner. In this case, the user and the agent interact via window displaying state transition diagram. When the user finishes scenario modification, the agent closes the window and restart from the modified state. To put it concretely, the agent can resolve deadlock. 2.1
Scenario Description
In this research, we use a method which separates an agent’s external behavior (interaction with other agents and environment) as a scenario from its internal mechanism[5]. A part of an agent which is to be a target of modification in rehearsal oriented testing becomes clear by separating scenarios from the agent’s internal model. Only scenarios are targets of modification in rehearsal oriented testing. Agents’ internal models are implemented by computer professionals in order to realize actions requested from scenarios on the simulation environment. On the other hand, since scenarios describe only agents’ external (and visible) behavior, users 1
http://www.ai.soc.i.kyoto-u.ac.jp/freewalk/
186
S. Yamane and T. Ishida
Fig. 2. Separate a scenario from an internal model
who are not computer professional can easily understand them. Users modify scenarios in order to work agents desirably for them. Figure 2 shows architecture which separates a scenario from an agent’s internal model. As shown in Figure 2, a scenario which describes an agent’s external behavior is input to a scenario interpreter and the scenario interpreter requests behavior of the agent. The agent internal model executes sensors and actuators according to the request. We use scenario description language Q to describe agent scenarios. Using Q, novice at using computer can describe agent scenarios based on extended finite state machine model. Detailed specification of Q is represented in [6]. 2.2
Processes of Scenario Rehearsal
We focused on two features of rehearsal in theater to realize rehearsal oriented testing. One is direction during performance. In rehearsal, actors perform in the same way as public performance and a director monitors them. If the director finds a problem in the actor’s performance, he/she directs the actor to fix the problem. Then he/she restarts performance to check whether the modification succeeds and to find another problem. Second feature is sharing of context. A director and actors share information such as a situation in a stage, other actors’ position and scripts. Based on these features, we construct rehearsal oriented testing methodology. Rehearsal oriented testing consists of two main processes. One is finding problems in a multiagent simulation, and the other is modifying the problems. Running a multiagent simulation and find a problem. The multiagent simulation runs and each agent behave in accordance with its scenario. This multiagent simulation should be visualized. A user monitors the visualized multiagent simulation and checks whether there are problems in the agents’ behavior. If the user finds a problem, he/she interrupts the multiagent simulation. Modifying a problem found. If a user finds a problem, he/she interrupts the simulation in order to modify the problem. The user tells the agent the way in which the user wants it to behave. The agent modifies its scenario based
An Agent Modeling Method Based on Scenario Rehearsal
187
on the user’s direction and context in the simulation. When the user finishes modification, he/she restarts the simulation. The first process will be mostly realized by visualization and faculties of a multiagent system. Therefore, in the rest of this paper, we focus on second process: how to modify problems in agent scenarios. 2.3
Setting of Modification Operators for Finite State Machine
In this paper, we use finite state machine model to describe agent scenarios. Finite state machine is easy to describe and understand by representing state transition diagram. This will help people who are not computer professionals to understand operation of scenarios. When a user designs interactions of agents, he/she first describes scenarios described on the basis of finite state machines. Then the user modifies them by rehearsal oriented testing. Therefore, modification operators of rehearsal oriented testing must include operators enough to modify a finite state machine. Typically, state transition model is defined as Q, C, A, q0 where Q is a set of states, C is a set of events can be observed by an agent, A is a set of actions can be executed by the agent, Δ is a set of state transition functions and q0 ∈ Q is an initial state. We express a condition of scenario execution st ∈ S at time t as st = Q, C, A, Δ, q0 , δt , qt . where δt is a state transition which is running at time t and qt is a state which a scenario is in at time t. Using this definition, scenario modification operators modify st . In other words, each element of st is considered as a target of modification. However, not all of the elements can be targets of the modification. C is a set of events which an agent can observe and A is a set of actions which an agent can execute in its environment. These are implemented as faculties of a multiagent system. Rehearsal oriented testing aims to refine agents’ behavior under given multiagent systems. For this reason, C and A can not be targets of the modification. As to q0 , since it is an initial state of the scenario, it seems unlikely that the user changes the initial state when the multiagent system is running. For this reason, q0 can not be a target of the modification. In addition, a user must not change δt and qt because they are the conditions of scenario execution. Excepting for these elements, the modification in rehearsal oriented testing targets Q and Δ. For Q and Δ of a scenario, possible modification operations are adding, modifying and deleting them. As to Q, since modifying a state means modifying state transitions which transit from and to the state, we consider only about adding and deleting it. From above reasons, there are following five modification operators. Adding a state. A user specifies a state to be added. Deleting a state. A user specifies a state to be deleted. Adding a state transition. A user specifies all element of a state transition (a current state, an event, an action and a next state). Modifying a state transition. A user specifies a state transition to be modified and how to modify each element of it.
188
S. Yamane and T. Ishida
Deleting a state transition. A user specifies a state transition to be deleted. In rehearsal oriented testing, a user gives an agent the information required for modification through dialog with the agent.
3
Interaction Protocol
In rehearsal oriented testing, a user modifies agent scenarios through dialog with agents using condition of scenarios and simulation environment at a certain moment in the simulation. However, it is very difficult to figure out a user’s intention from completely free interaction. Therefore we have to give constraints on their interaction. In addition, a user can understand what he/she should do to modify its scenario by following the interaction protocol. In this section, we explain about the protocol. This interaction protocols should be natural and simple as rehearsal in order to reduce cost of rehearsal oriented testing. We use context in a simulation to make dialog simple and natural interaction between a user and agents. In concrete terms, if a user starts modification in time t, qt ∈ Q, δt ∈ Δ of agent scenarios can be used as the context. Previously, we explained that targets of modification by rehearsal oriented testing are states and transition functions of agent scenarios. A user has to specify a state to be deleted in deleting state, current state and next state in adding state transition, current state and next state of transition function to be modified in modifying state transition and transition function to be deleted in deleting state transition respectively. Dialog between a user and an agent can be more simple using qt , δt in these modification cases. Let qc be a state which a user is modifying and let δc be a state transition which a user is modifying in rehearsal oriented testing. At the time t when a user starts the modification, qt is assigned to qc and δt is assigned to δc . For deleting state, qc is used for deleted state. For adding state transition, qc is used for current state. For modifying and deleting state, δc is used for state transition to be modified or deleted. We design interaction protocol taking into account the context explained above. At the moment when a user starts modification, scenario execution condition can be classified into two. One is an agent is executing an action in a transition function. The other is an agent is waiting some events in a certain state. We explain interactions which start from each condition. Figure 3 shows the interaction protocol when a modification starts. As explained above, at the moment when the modification starts, an agent is executing action in a certain state transition or waiting events in a certain state. The interaction is first divided into two according to which condition an agent is in. If the agent is executing an action, a user can modify a state transition which includes the action. In this case, the state transition is δc . When the user finishes modifying the state transition or the user does not want to modify the state transition, the user can move to modification of a state. When the user move to modification of the state, the current state of the state transition δc is assigned to qc .
An Agent Modeling Method Based on Scenario Rehearsal
189
Fig. 3. Interaction protocol when the modification starts
Fig. 4. Interaction protocol for modifying and deleting a state transition
If the agent is waiting events in a certain state, the user starts interaction to modify the state. The state is assigned to qc . The user can add a state, delete a state and add a state transition in this case. When the user wants to modify or delete a state transition, he/she specify the state transition and move to interaction to modifying the state transition assigning the state transition to δc . 3.1
Interaction Protocol in Executing State Transition
Figure 4 shows an interaction when an agent is executing state transition. A user can modify and delete a state transition in this protocol. Modification of a state transition. A user first selects an event, an action or a next state of the state transition to be modified. Then the user gives instruction on how to modify it. The user repeats this interaction if he/she wants to modify more. Deletion of a state transition. A user can delete δc by telling the agent to delete the state transition. When a user wants to modify state such as adding a state transition, he/she can tell the agent to modify a state. In this case, a current state of the state transition δc is assigned to qc .
190
S. Yamane and T. Ishida
Fig. 5. Interaction protocol for adding and deleting states and adding state transitions
3.2
Interaction Protocol in Waiting Events
Figure 5 shows the interaction protocol while waiting events in a certain state. A user can add/delete a state and add a state transition. Addition of a state. A user tells the agent to add a state then the agent creates a new state. Deletion of a state. A user specifies a state to be deleted and the agent deletes it. Addition of a state transition. First, qc is selected as a current state. Then an agent asks a user information required for a state transition. In the interaction protocol shown in Figure 5, the user first specifies an event to be observed. And then the user specifies an action and at last the user specifies a next state. If a user wants to modify or delete a state transition which has qc as a current state, the user specifies the state transition and moves to interaction to modify the state transition which is shown in Figure 4.
4
Example
We show an example of rehearsal oriented testing using an evacuation simulation. The simulation is performed as follows. There are dozens of agents in a room. The agents start escaping from the room when they hear sound of a siren. This simulation finishes when all the agents go out from the room. A user monitors this simulation and checks if all the agents can go out. Let us suppose a situation that one agent does not move while the other agents are heading to an exit. First, a user analyzes what is happening through visualized simulation. Visualized simulation is shown in Figure 6. The user finds that the agent is facing a wall. Then the user speaks to the agent to know what the agent is doing. The agent shows its scenario in response and says “Now I am in searching state in this scenario.” Figure 7 show the scenario. In this case, the scenario is in searching state. The agent is looking for another agent or an
An Agent Modeling Method Based on Scenario Rehearsal
191
Fig. 6. Evacuation simulation
Fig. 7. Initial scenario
exit in this state. From this scenario and visualized agent which is facing the wall, the usercan see the agent does not move since it can find nothing. Next, the user modifies this problem based on the analysis above. The agent should turn around if it can not find anything in searching state. Therefore the user and the agent talk in order to modify scenario as follows. When the dialog starts, a state of the interaction protocol is Modifying a state state in Figure 5. The user request to add an action when the agent can not find anything saying “Please add an action when you can do nothing in a few seconds.” Then the agent says “What should I do when I can do nothing in a few seconds?” and the interaction protocol transits to Adding an action state. The user says “Please turn around.” and the agent replies “Please specify a next state.” according to the interaction protocol. The user says “searching state.” in order for the agent to back to search turning around. Modified scenario in this way is shown in Figure 8. The new state transition in searching state was added. This example shows how to find and modify a problem of a scenario by rehearsal oriented testing. Visualized simulation and information about a scenario from an agent are available for a user. The user identify a problem from these information and decide how to modify it. In addition, dialog based on the
192
S. Yamane and T. Ishida
Fig. 8. Modified scenario
interaction protocol helps a user to know what an agent needs to modify the scenario properly.
5
Application
In this paper we proposed rehearsal oriented testing in order to modify agent scenarios for multiagent simulation. However, we believe there are many other applications using rehearsal oriented testing. Rehearsal oriented testing can be applied to control of robots. In order to do this, we have to activate a robot using an agent system and install interfaces required to talk with humans such as a speaker and a microphone. Furthermore, implementing faculties explained in this paper, we can perform rehearsal oriented testing to the robot. For example, by saying “When I say ‘Please fetch a newspaper’ to you, go outside...” to the robot, a user can extend the robot’s behavior. In addition, if there is no newspaper outside, the robot can do nothing because it does not know what it should do. Then the user can modify the robot’s behavior by saying “If there is no newspaper...” to the robot. Rehearsal oriented testing includes a process in which a user talks to an agent to know state of an agent scenario. Taking advantage of this, when an interesting event occurs in a simulation, rehearsal oriented testing can be used as a tool to inspect what is happening in the simulation. In order to realize this, rehearsal oriented testing should be able to target more than one agent at the same time. In addition, rehearsal oriented testing should show the user not only a scenario execution condition of the agent, but also history of interaction with other agents and environment.
6
Conclusion and Future Work
In this paper, we proposed rehearsal oriented testing in order for novice at using computer to modify agent scenarios. In particular, we focused on interaction protocol between a user and an agent under which a user can modify his/her agent scenario through the interaction.
An Agent Modeling Method Based on Scenario Rehearsal
193
We accomplished the following two goals to realize rehearsal oriented testing. Setting of modification operators. We set the modification operators for rehearsal oriented testing. If scenarios are described on the basis of finite state machines, they are addition of a state, deletion of a state, addition of a state transition, modification of a state transition and deletion of a state transition. Design of an interaction protocol. Under this protocol, an agent tells a user what kind of directions the user can give in order for the user to know modification procedure. In addition, this protocol realizes more natural dialog using context in a simulation. Rehearsal oriented testing is one process of development of a multiagent system. Development tools and methodologies for multiagent systems have been studied for years[7][8]. In contrast to these researches, our research aims to help users who are not computer professionals to make agents behave properly. Users, such as participants of participatory simulations and users of service agents, have only to consider interaction scenarios of agents and the agents which behave according to the scenarios are implemented by computer professionals. In rehearsal oriented testing, a human finds and modifies agent scenarios based on visualized behavior of agents. This is effective if the user wants to modify behavior of one agent or simple interaction such as one-on-one interaction. However, it is difficult to find out a problem from many scenarios which interact with each other. For this purpose, techniques which help us to analyze interaction between the agents such as those implemented in [9] are required. Our future works also include learning of directions from a user in rehearsal oriented testing. Rehearsal oriented testing is a method to modify and extend agent models which are designed from some sort of techniques such as participatory simulations. However, different from theaters in which limited situations can exist, there can exist enormous situations in a multiagent simulation. It is not realistic to simulate immense situations in the simulation and perform rehearsal oriented testing for each situation. Therefore, we have to make agents which not only follow a user’s directions but also learn them. The agents deduce the user’s intention based on the learning and updates its scenario autonomously. If this mechanism is realized, we can make agents which adapt its scenario to a situation which is not considered in designing the agents and a change of their environment.
Acknowledgment This research was partially supported by a Grant-in-Aid for Scientific Research (A) (18200009, 2006-2008) from Japan Society for the Promotion of Science (JSPS) and a Grant-in-Aid for JSPS Fellows.
194
S. Yamane and T. Ishida
References 1. Drogoul, A., Vanbergue, D., Meurisse, T.: Multi-agent Based Simulation: Where Are the Agents? In: Sichman, J.S., Bousquet, F., Davidsson, P. (eds.) MABS 2002. LNCS, vol. 2581, pp. 1–15. Springer, Heidelberg (2003) 2. Guyot, P., Drogoul, A., Lemaitre, C.: Using emergence in participatory simulations to design multi-agent systems. In: Proceedings of The Fourth International Joint Conference on Autonomous Agents and Multi-Agent Systems (AAMAS) (2005) 3. Murakami, Y., Sugimoto, Y., Ishida, T.: Modeling human behavior for virtual training systems. In: The Twentieth National Conference on Artificial Intelligence (AAAI 2005) (2005) 4. Nakanishi, H., Ishida, T.: FreeWalk/Q: Social interaction platform in virtual space. In: ACM Symposium on Virtual Reality Software and Technology (VRST 2004), pp. 97–104 (2004) 5. Ishida, T.: Q: A scenario description language for interactive agents. IEEE Computer 35(11), 42–47 (2002) 6. Ishida, T., Yamane, S.: Introduction to scenario description language Q. In: International Conference on Informatics Research for Development of Knowledge Society Infrastructure (ICKS 2007). IEEE Computer Society, Los Alamitos (2007) 7. Wooldridge, M., Jennings, N.R., Kinny, D.: The Gaia methodology for agentoriented analysis and design. Autonomous Agents and Multi-Agent Systems 3(3), 285–312 (2000) 8. Nwana, H.S., Ndumu, D.T., Lee, L.C., Collis, J.C.: ZEUS: a toolkit and approach for building distributed multi-agent systems. In: Etzioni, O., M¨ uller, J.P., Bradshaw, J.M. (eds.) Proceedings of the Third International Conference on Autonomous Agents (Agents 1999), pp. 360–361 (1999) 9. Ndumu, D.T., Nwana, H.S., Lee, L.C., Collis, J.C.: Visualising and debugging distributed multi-agent systems. In: Proceedings of the Third International Conference on Autonomous Agents (Agents 1999), pp. 326–333 (1999)
Fast Partial Reallocation in Combinatorial Auctions for Iterative Resource Allocation Naoki Fukuta1 and Takayuki Ito2 1
2
Shizuoka University, Hamamatsu Shizuoka 4328011, Japan [email protected] http://whitebear.cs.inf.shizuoka.ac.jp/ Nagoya Institute of Technology, Gokiso-cho Nagoya 4668555, Japan
Abstract. In this paper, we propose enhanced approximation algorithms of combinatorial auction that are suitable for the purpose of periodical reallocation of items. Our algorithms are designed to effectively reuse the last solutions to speeding up initial approximation performance. We show experimental results that show our proposed algorithms outperform existing algorithms in some aspects when the existing bids are not deleted. Also, we propose an enhanced algorithm that effectively avoids undesirable reuse of last solutions in the algorithm. This is especially effective when some existing bids are deleted from the last cycle. Furthermore, our algorithms satisfy two desirable properties: WPM for final results and Weak-WPM for intermediate results.
1
Introduction
Combinatorial auctions, one of the most popular market mechanisms, have a huge effect on electronic markets and political strategies. For example, Sandholm et al. [1] proposed a market using their innovative combinatorial auction algorithms. It is well-known that the FCC tried to employ combinatorial auction mechanisms for assigning spectrums to companies [2]. Resource allocation for agents on ubiquitous computing environment is a good example to understand the needs of short time approximation of combinatorial auctions. On ubiquitous computing environment, there are typically limited amount of resources(sensors, devices, etc.) that do not cover all needs for all agents. Due to some reasons (physical limitations, privacy reasons, etc.), most of the resources cannot be shared with other agents. Furthermore, agents will use two or more resources at a time to realize desirable services for users. Of course each agent provides services to its own user, agent may be self-interested. Therefore, combinatorial auction mechanism is suitable for such situation since it is well-known mechanism for effective resource allocation to self-interested agents. Here, consider there are 256 resources, 100 agents, and each agent has 100 to 200 bids, there will be 10,000 to 20,000 bids for 256 goods in an auction. On ubiquitous computing scenario, physical location of users are always changing, resources should be reallocated in a certain period to catch up those changes. For better usability, the time period of resource reallocation will be 0.1 to several A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 195–206, 2009. c Springer-Verlag Berlin Heidelberg 2009
196
N. Fukuta and T. Ito
seconds depending on services provided there. Therefore, it is strongly demanded to get winners of an auction within such extremely short time period. In general, the optimal winner determination problem of combinatorial auction is NP-hard[3]. Thus, much work focuses on tackling the computational costs for winner determination [4][3][1]. These works basically try to achieve the optimal solution in winner determination. Although these works can be used for attaining approximated solutions and they perform well in certain conditions, there remain possibilities for further improvements focused on approximation. Some works [5] [6] [7] try to achieve approximate solutions in winner determination. A recent well-known theoretical work is Lehmann et al. [5]. In [5], they proposed a greedy approximation algorithm that relaxes economic efficiency and proves the truthfulness in a certain condition. The paper [5] rather focuses on the more elegant theoretical aspects, i.e., giving truthfulness for an auction mechanism with approximated allocations of items. However, it is well known that the performance of [5] is really nice that can successfully handle a very large number of bids and goods in an auction. Also, it is well known that the approximation algorithms presented in [6] and [7] generally perform very well even if the number of bids is relatively large. Recently, we proposed a set of extended approximation algorithms that are based on Lehmann’s approach. In [8], we showed that our algorithm perform well when the number of bids is very large and there is hard time constraint. Also we reported detailed comparisons of our algorithms and existing algorithms(e.g., [9]). However, since those existing algorithms are ‘offline’ algorithms, we need to recalculate the winners when some bids are added to or deleted from the auction even when the modification of bids is little. It will be helpful when the most recent results can be reused to the approximation process of the current auction. However, those algorithms did not consider about reusing past approximated results for performance improvement. In this paper, we propose enhanced approximation algorithms of combinatorial auction that are suitable for the purpose of periodical reallocation of items we mentioned above. This paper is organized as follows. In section 2, we present some necessary backgrounds about our research, including definition of the problem and existing algorithms. Section 3 describes our proposed algorithms that are designed to effectively reuse the last results to speeding up initial approximation performance after the modification of bids in the auction. In Section 4, We show experimental results that shows our proposed algorithms outperform existing algorithms in some aspects when the existing bids are not deleted. Also, we report the effectiveness of an enhanced algorithm that effectively avoid undesirable reuse of last solutions in the algorithm. In Section 5, we show our algorithms satisfy two desirable properties: WPM for final results and Weak-WPM for intermediate results. In Section 6, we show related works to clarify our contributions in this paper. Finally, section 7 concludes the paper.
Fast Partial Reallocation in Combinatorial Auctions
2 2.1
197
Preliminaries Combinatorial Auction
Combinatorial auction is an auction mechanism that allows bidders to locate bids for a bundle of items rather than single item. The winner determination problem is defined as follows[3] : The set of bidders is denoted by N = 1, . . . , n, and the set of items by M = {m1 , . . . , mk }. |M | = k. Bundle S is a set of items : S ⊆ M . We denote by vi (S), bidder i’s combinatorial bid for bundle S. An allocation of the items is described by variables xi (S) ∈ {0, 1}, where xi (S) = 1 if and only if bidder i wins the bundle S. An allocation, xi (S), is feasible if it allocates no item more than once, xi (S) ≤ 1 i∈N Sj
for all j ∈ M . The winner determination problem is the problem to maximize total revenue vi (S)xi (S) max X
i∈N,S⊆M
for feasible allocations X xi (S) . Throughout this paper, we only consider the auctions that are single-sided, with a single seller and multiple buyers to keep the simplicity of discussions. It can be extended to the reverse situation with a single buyer and multiple sellers, and the two-sided cases[10]. 2.2
Lehmann’s Greedy Winner Determination
Lehmann’s greedy algorithm [5] is a very simple but powerful linear algorithm for winner determination on combinatorial auctions. Here, a bidder declaring < a, s >, with s ⊆ M and a ∈ R+ will be said to put out a bid b =< a, s >. Two bids b =< a, s > and b =< a , s > conflict iff s ∩ s = ∅. The greedy algorithm can be described as follows: (1)The bids are sorted by some criterion. The paper[5] proposed sorting the list L by descending average amount per good. More generally, they proposed sorting L by a criterion of the form a/|s|c for some number c, c ≥ 0, possibly depends on the number of goods, k. (2) A greedy algorithm generates an allocation. L is the sorted list in the first phase. Walk down the list L, accepting bids if the goods demanded are still unallocated and not conflicted. In the paper[5], they argued that c = 1/2 is the best parameter for approximation when the norm of worst case performance is considered. Also they have shown that the mechanism is truthful when single-minded bidders are assumed and their proposed pricing scheme is used.
198
N. Fukuta and T. Ito
2.3
Hill-Climbing Search
In [8], we have shown that hill-climbing approach performs well when an auction has a massively large number of bids. In this section, we summarize our proposed algorithms. Lehmann’s greedy winner determination could succeed specifying the lower bound of the optimality in its allocation[5]. The straightforward extension of the greedy algorithm is to construct a local search algorithm that continuously updates the allocation so that the optimality is increased. Intuitively, one allocation corresponds to one state of a local search. The inputs are Alloc and L. L is the bid list of an auction. Alloc is the initial greedy allocation of items for the bid list. 1: function LocalSearch(Alloc, L) 2: RemainBids:= L - Alloc; 3: for each b ∈ RemainBids as sorted order 4: if b conflicts Alloc then 5: Conf licted:=Alloc - consistentBids({b}, Alloc); 6: N ewAlloc:= Alloc - Conf licted + {b}; 7: ConsBids:= 8: consistentBids(N ewAlloc, RemainBids); 9: N ewAlloc:=N ewAlloc+ConsBids; 10: if price(Alloc) < price(N ewAlloc) then 11: return LocalSearch(N ewAlloc,L); 12: end for each 13: return Alloc
The function consistentBids finds consistent bids for the set N ewAllocation by walking down the list RemainBids. Here, a new inserted bid will wipe out some bids that are conflicted with the inserted bid. So there will be free items to allocate after the insertion. The function consistentBids tries to insert the other bids for selling the higher value goods as many as possible. 2.4
Local Search for Multiple Values of the Sorting Criterion c
The optimality of allocations got by Lehmann’s algorithm (and the followed hill-climbing) deeply depends on which value was set to the bid sorting criterion c. Lehmann et al. reported c = 1/2 is the approximately best value. But the optimal values for each auction are varied from 0 to 1 even if the number of items is constant. Here, we use an enhancement for our local search algorithm with parallel search for the sorting criterion c. In the algorithm, the value of c for Lehmann’s algorithm is selected from a pre-defined list. It is reasonable to select c from neighbors of 1/2, namely, C = {0.0, 0.1, . . . , 1.0}. The results are aggregated and the best one (that has highest revenue) is selected as the final result.
Fast Partial Reallocation in Combinatorial Auctions
2.5
199
Other Approximation Approaches
Zurel and Nisan[6] proposed a very competitive approximate winner determination algorithm for combinatorial auctions. The main idea is a combination of approximated positive linear program algorithm for determining initial allocation and stepwise random updates of allocations. Hoos[7] proposed that a generic random walk SAT solver may perform well for approximation of combinatorial auctions. In [7], Casanova algorithm has been proposed for the purpose of approximating winner determination. It is based on scoring each search state using the revenue-per-good of the corresponding allocation.
3
Enhanced Approximation Algorithms
3.1
Fast Partial Reallocation by Using Last Result
In the setting of periodical resource re-allocation scenario, winner determination occurs when some bids are revised. Theoretically, we need to recalculate winners even if only one bid is changed in the auction. However, in some cases, it is useful to re-use winners of previous auction when the change is small so that it has small effects to the next winner determination process. Following simple algorithm reuses the approximation result in last cycle when recalculation is needed due to change of bids in the auction. Here, we assume winners of the last cycle (LastW inners) and bidset of the last cycle (LastBids) are known. 1: Function PartialReallocationA(LastBids,LastW inners,CurrentBids) 2: AddedBids := CurrentBids - (LastBids ∩ CurrentBids); 3: DeletedBids := LastBids - (LastBids ∩ CurrentBids); 4: W inners := LastW inners; 5: foreach d ∈ DeletedBids 6: if d ∈ W inners then 7: W inners := W inners − d; 8: foreach a ∈ AddedBids 9: foreach w ∈ W inners 10: if b(w) = b(a) and v(w) < v(a) then 11: W inners := W inners − w + a; 12: W inners := LocalSearch(W inners,CurrentBids); 13: return W inners
At the first step, the algorithm deletes winners that no longer valid due to deletion of bids. And then, some winner bids are replaced by newly added bids. Note that we only replace when the bundle of bids are just equal to avoid ordering problem of newly added bids. Modification of a bid through cycles will be treated as a combined operation of deletion of previous bid and addition of the renewed bid.
200
3.2
N. Fukuta and T. Ito
Eliminating Bad Reallocations within Short Overhead
Generally speaking, performance of reusing partial result of similar problem depends on the problem. Therefore, in some cases, reuse of last result may cause performance decreases. To avoid such a situation, we slightly modified our algorithm to switch initial allocation by evaluating its performance. Here, the modified algorithm simply compares the reused result with greedy allocation. Then, the better one is used for the seed of Hill-climbing improvement. Note that both our reallocation algorithm and greedy allocation algorithm complete their executions in very short time. Therefore, the computational overhead is expected to be negligible. 1: Function PartialReallocationX(LastBids,LastW inners,CurrentBids) 2: AddedBids := CurrentBids - (LastBids ∩ CurrentBids); 3: DeletedBids := LastBids - (LastBids ∩ CurrentBids); 4: W inners := LastW inners; 5: foreach d ∈ DeletedBids 6: if d ∈ W inners then 7: W inners := W inners − d; 8: foreach a ∈ AddedBids 9: foreach w ∈ W inners 10: if b(w) = b(a) and v(w) < v(a) then 11: W inners := W inners − w + a; 12: GreedyW inners := GreedySearch(CurrentBids); 13: if price(W inners) =< price(GreedyW inners) then 14: W inners := GreedyW inners; 15: W inners := LocalSearch(W inners,CurrentBids); 16: return W inners
4
Experimental Results
We implemented our algorithms in a C program for the following experiments. We also implemented the Casanova algorithm in a C program. However, for the following experiments, we used Zurel’s C++ based implementation that are shown in [6]. The experiments were done with above implementations to examine the performance differences among algorithms. The program was employed on our Mac with Mac OS X 10.4, CoreDuo 2.0GHz CPU, and 2GBytes of memory. We conducted several experiments. In each experiment, we compared the following search algorithms. “greedy(C=0.5)” uses Lehmann’s greedy allocation algorithm with parameter (c = 0.5). “greedy-3” uses the best results of Lehmann’s greedy allocation algorithm with parameter (0 ≤ c ≤ 1 in 0.5 steps). “HC(c=0.5)” uses a local search in which the initial allocation is Lehmann’s allocation with c = 0.5 and conduct the hill-climbing search shown in the previous section. “HC-3” uses the best results of the hill-climbing search with parameter
Fast Partial Reallocation in Combinatorial Auctions
201
(0 ≤ c ≤ 1 in 0.5 steps). Also, we denote Casanova algorithm as “casanova” and Zurel’s algorithm as “Zurel”. In the following experiments, we used 0.2 for epsilon value of Zurel’s algorithm in our experiments. This value appears in [6]. Also we used 0.5 for np and 0.15 for wp on Casanova that appear in [7]. Note that we set maxT rial to 1 but maxSteps to ten times of the number of bids in the auction. We conducted detailed comparisons among our past presented algorithms and other existing algorithms mentioned above. The detail of the comparisons has shown in [9] and [11]. In [9] and [11], we prepared datasets having 20,000 bids in an auction. The datasets were produced by CATS[12] with default parameters in 5 different distributions. It contains 100 trials for each distribution. Each trial is an auction problem with 256 items and 20,000 bids1 . However, since CATS common datasets provide only static bids for an auction, we prepared an extended usage for those datasets to include dynamic changes of bids in an auction. Procedure: In each auction, the bid set is divided into 10 blocks by the order of bid generation (i.e. bid id). The bid set is modified totally 10 times and the modification will be done in each 1 second. In each period of 1 second, a block is marked as hidden so that bids within these marked blocks are treated as ‘deleted bids’. For example, at first period, the first block is marked as hidden so left bids (second to 10th blocks) are used for winner determination. After 1 second, the mark is moved to second block (i.e. first, and third to 10th blocks are used) and the winner determination process is restarted due to this change. Here, we can see it as the bids in first block are newly added to the auction and the bids in second block are deleted from the auction. This process is repeated until the mark has been moved to the 10th block. And finally, all marks are cleared and the winner determination process is restarted with full bids in the auction. Ordinary algorithms should be completely restarted in each cycle. However, when we use our proposed reallocation algorithms, some intermediate results can be reused in the next cycle in the same auction. Since the bid set in last period is completely equal to the bids of the auction, the result last periods can be compared to our previous experimental results. Table 1 shows the experimental result on the dataset with 20,000 bids in an auction focused on the execution time of short time approximation. Due to difficulty of attaining optimal values, we normalized all values as Zurel’s results equal 1. The name of each distribution is referred from [12]. We prepared cut-off results of Casanova and HC. For example, “casanova-10ms” denotes the result of Casanova within 10 milliseconds. Also we prepared a variant of our algorithm that has a suffix of ‘-seq’ or ‘-para’. The suffix ‘-seq’ denotes the algorithm is completely executed in sequential that is equal to be executed on a single CPU computer. For example, greedy-3-seq denotes that the execution time is just the sum of execution times of three threads. The suffix ‘-para’ denotes the algorithm 1
Due to difficulty of preparing dataset, we only prepared 5 distributions. Producing dataset with other distributions is difficult in feasible time. More detailed experimental results are shown in [9].
202
N. Fukuta and T. Ito Table 1. Time Performance on 20,000bids-256items
is completely executed in parallel manner, the three independent threads are completely executed in parallel. Here, we used ideal value for ‘-para’ since our computer has only two core in the CPU. The actual execution performance will be between ‘-seq’ and ‘-para’. Additionally, we added results with the name ‘AHC-’ or ‘XHC-’ in the same table. They are the average approximated results of the last periods of auctions with our proposed algorithms PartialReallocationA and PartialReallocationX, respectively2 . In most distributions, Zurel-1st takes more than 1 second but the obtained optimality is lower than greedy-3-seq. However, our proposed ‘HC-3’ performs better or slightly lower although their computation times are shorter than Zurel1st and Zurel, excluding L3. Surprisingly, in most cases, the results of ‘XHC3-seq-100ms’ are better than ‘HC-3-seq-1000ms’ while their spend computation time are only 1/10. This fact shows that our ‘XHC-3’ could reuse approximated results of previous cycles effectively. Figure 1 shows the average performance curves of ‘HC-3-para’, ‘AHC-3-para’, and ‘XHC-3-para’. In Figure 1, average performances of algorithms in each distribution are separately plotted. Since it can be viewed that each 1-sec period has its own auction problem, the optimal total revenue cannot be a fixed value in the whole period. For this reason, the plotted performances are normalized as the result of HC-3-para-1000msec be 1 in each 1-sec period. For example, when the value in the time 2100msec was 0.95, it means that the total revenue of approximated result was 95 percent of the maximum approximated value 2
We omitted the results of L2 for ‘AHC’ and ‘XHC’ since the obtained results of HC(c=0.5)-100ms are good enough and it is easily predicted to get the same performance by our ‘AHC’ and ‘XHC’.
Fast Partial Reallocation in Combinatorial Auctions
203
1.01 1 1.005
1 Normalized Performance
Normalized Performance
0.98
0.96
0.94
0.995
0.99
0.985
0.92 0.98 0.9
0.975 L3-HC-3-para L3-AHC-3-para L3-XHC-3-para
L4-HC-3-para L4-AHC-3-para L4-XHC-3-para
0.88
0.97 0
2000
4000
6000
8000
10000
0
2000
4000
Elapsed Time
6000
8000
10000
Elapsed Time
1.01
1.05
1
1
0.95 Normalized Performance
Normalized Performance
0.99
0.98
0.97
0.9
0.85
0.8
0.96 0.75 0.95
0.7 L6-HC-3-para L6-AHC-3-para L6-XHC-3-para
0.94
L7-HC-3-para L7-AHC-3-para L7-XHC-3-para 0.65
0
2000
4000
6000 Elapsed Time
8000
10000
0
2000
4000
6000 Elapsed Time
8000
10000
Fig. 1. Perfromance Changes on 10-fold periodical bid replacement
attained within the period starting from 2000msec to 3000msec by ‘HC-3-para’. Therefore we can see how each algorithm can improve the approximated results in each cycle. On L7, both ‘AHC-3’ and ‘XHC-3’ perform very well compared to ‘HC-3’. However, on L3, L4, and L6, the performance of ‘AHC-3’ is far worse than ‘HC-3’ excluding the final period. This is because the deletion and addition of bids in the same time will cause a mis-choice of initial allocations in each cycle when there is no ‘bad reallocation avoiding mechanism’ within the algorithm. In contrast, ‘XHC-3’ only reuses the results of last cycle when it is better than the vanilla greedy result. Actually, the results of ‘XHC-3’ are constantly better than ‘AHC-3’ and there is no performance decrease compared to ‘HC-3’ in any time of the period.
5
Discussion
In real world, a loser who might be a winner in the optimal allocation could know the winner’s bidding price in an approximate allocation after the auction finishes. This loser might place a higher price than the winner’s even if they placed bids on the same bundle. This could result in ”unfair” allocations for bidders. Therefore, we proposed two desirable properties, Winner-Price-Monotonicity and WeakPrice-Monotonicity to avoid ‘unfair’ allocations[11].
204
N. Fukuta and T. Ito
Definition 1. (Winner-Price-Monotonicity: WPM) For two bundles B and B , if B ⊆ B and vi (B) > vj (B ), then j must not win bundle B . Definition 2. (Weak-Winner-Price-Monotonicity: Weak-WPM) For one bundle B, if vi (B) > vj (B), then j must not win bundle B. We gave proofs about following propositions in [11]. Proposition 1. Our proposed determination algorithms, except for the simulated annealing-based algorithm, produce allocation Wf in that satisfies WPM when the algorithm reaches an end. Proposition 2. In terms of any allocations that are achieved during computation (as an anytime algorithm), our proposed winner determination algorithms, except for the simulated annealing-based algorithm, satisfy Weak-WPM. We designed our newly proposed algorithms (‘AHC-3’ and ‘XHC-3’) to keep these desirable properties. Here, we show that above propositions can also be applied to the proposed reallocation algorithms. For this purpose, we give Lemma 1 and its proof below. Lemma 1. Our proposed initial winner set reallocation algorithm satisfies WeakWPM. Proof for Lemma 1. Consider about function PartialReallocationA, when LastBids = LastW inners = ∅, it is equal to simple Lehmann’s greedy allocation. Therefore, initial allocation for initial cycle (LastBids = LastW inners = ∅) satisfies WPM. In function PartialReallocationA, DeletedBids will be deleted from W inners when they are included in last winners (LastW inners). Here, recall the definitions of WPM and Weak-WPM, there is no effect about deleting bids from winners. A special example about it is that when W inners = ∅, it satisfies WPM and Weak-WPM since they only consider about invalid winners. Therefore, the deletion mechanism will not take effect about WPM and Weak-WPM. Also, in function PartialReallocationA, last winners (W inners) are replaced by most higher one when bids for just same bundle are newly added. Here, it is guaranteed that all winner bids in W inners are placing highest prices for the bundles. Therefore, there is no bid b (x) ∈ W inners such that v (x) < v(X) ∈ CurrentBids. Note that some bids will be added to W inners such that they do not conflict to the current bids in W inners. Here, since we used Lehmann’s greedy algorithm for this purpose, and the fact that the procedure for filling bids for non-allocated bundles will not break Weak-WPM[11], this will not break Weak-WPM. From above, any initial winner set generated by our reallocation algorithm satisfies Weak-WPM. Also any initial winner set generated in function PartialReallocationX satisfies Weak-WPM. Since it is same winner set in function PartialReallocationA or attained by simple Lehmann’s greedy allocation, and both satisfies at least Weak-WPM. 2
Fast Partial Reallocation in Combinatorial Auctions
205
Recall the proof in [11] and the proof about above lemma 1, it is clear that proposition 2 is satisfied for our proposed reallocation algorithms. Also, it is easy to extend the proof in [11] to the proof of proposition 1 for our proposed reallocation algorithms3 .
6
Related Work
There have been a lot of works on the optimal algorithm for winner determination in combinatorial auctions[13]. Recently, Dobzinski et,al. proposed improved approximation algorithms for auctions with submodular bidders[14]. Lavi et,al, reported an LP based algorithm that can be extended to support the classic VCG[15]. Those researches are mainly focused on theoretical aspects. In the contrast to those papers, we rather focus on experimental and implementation aspects. Those papers did not present experimental analysis about the settings with large number of bids we presented in this paper. Also, Guo[16] proposed local-search based algorithms for large number of bids in combinatorial auction problems. However, their approach needs to adjust some parameters to make the algorithms perform satisfactory that is not necessary in our algorithms. CPLEX is a well-known, very fast linear programming system. In [6], Zurel et, al. evaluated the performance of their presented algorithm with many data sets, compared with CPLEX, and other existing implementations. While the version of CPLEX used in [6] is not up-to-date, the shown performance of Zurel’s algorithm is approximately 10 to 100 times faster than CPLEX. Therefore, the performance of Zurel’s and ours shown in this paper will be competitive enough to CPLEX and similar systems. Above approaches are based on offline algorithms and there are no considerations about addition and deletion of bids in their approximation processes. Although our algorithms are not strict online algorithms, it is possible to reuse the last results when bids are modified and recalculation is necessary.
7
Conclusions
In this paper, we proposed enhanced approximation algorithms of combinatorial auction that are suitable for the purpose of periodical reallocation of items. Our proposed algorithms effectively reuse the last solutions to speeding up initial approximation performance. The experimental results showed that our proposed algorithms outperform existing algorithms in some aspects when the existing bids are not deleted. However, we found that in some cases reuse of last solutions may cause worse performance compared to ordinary approximation from scratch. We proposed an enhanced algorithm that effectively avoids undesirable reuse of last solutions in the algorithm. We showed this is especially effective when some existing bids are deleted from the last cycle. Furthermore, We showed that our proposed algorithms satisfy two desirable properties: WPM for final results and Weak-WPM for intermediate results. 3
Due to space limitation, detailed proofs are omitted.
206
N. Fukuta and T. Ito
References 1. Sandholm, T., Suri, S., Gilpin, A., Levine, D.: Cabob: A fast optimal algorithm for winner determination in combinatorial auctions. Management Science 51(3), 374–390 (2005) 2. McMillan, J.: Selling spectrum rights. The Journal of Economic Perspectives (1994) 3. Cramton, P., Shoham, Y., Steinberg, R.: Combinatorial Auctions. MIT Press, Cambridge (2006) 4. Fujishima, Y., Leyton-Brown, K., Shoham, Y.: Taming the computational complexity of combinatorial auctions: Optimal and approximate approarches. In: Proc. of the 16th International Joint Conference on Artificial Intelligence (IJCAI 1999), pp. 548–553 (1999) 5. Lehmann, D., O’Callaghan, L.I., Shoham, Y.: Truth revelation in rapid, approximately efficient combinatorial auctions. Journal of the ACM 49, 577–602 (2002) 6. Zurel, E., Nisan, N.: An efficient approximate allocation algorithm for combinatorial auctions. In: Proc. of the Third ACM Conference on Electronic Commerce (EC 2001), pp. 125–136 (2001) 7. Hoos, H.H., Boutilier, C.: Solving combinatorial auctions using stochastic local search. In: Proc. of the AAAI 2000, pp. 22–29 (2000) 8. Fukuta, N., Ito, T.: Towards better approximation of winner determination for combinatorial auctions with large number of bids. In: Proc. of The 2006 WIC/IEEE/ACM International Conference on Intelligent Agent Technology(IAT 2006), pp. 618–621 (2006) 9. Fukuta, N., Ito, T.: Short-time approximation on combinatorial auctions – a comparison on approximated winner determination algorithms. In: Proc. of The 3rd International Workshop on Data Engineering Issues in E-Commerce and Services (DEECS 2007), pp. 42–55 (June 2007) 10. Parkes, D.C., Cavallo, R., Elprin, N., Juda, A., Lahaie, S., Lubin, B., Michael, L., Shneidman, J., Sultan, H.: Ice: An iterative combinatorial exchange. In: The Proc. 6th ACM Conf. on Electronic Commerce (EC 2005) (2005) 11. Fukuta, N., Ito, T.: Periodical resource allocation using approximated combinatorial auctions. In: Proc. of The 2007 WIC/IEEE/ACM International Conference on Intelligent Agent Technology (IAT 2007) (2007) 12. Leyton-Brown, K., Pearson, M., Shoham, Y.: Towards a universal test suite for combinatorial auction algorithms. In: Proc. of EC 2000, pp. 66–76 (2000) 13. de Vries, S., Vohra, R.V.: Combinatorial auctions: A survey. International Transactions in Operational Research 15(3), 284–309 (2003) 14. Dobzinski, S., Schapira, M.: An improved approximation algorithm for combinatorial auctions with submodular bidders. In: SODA 2006: Proceedings of the seventeenth annual ACM-SIAM symposium on Discrete algorithm, pp. 1064–1073. ACM Press, New York (2006) 15. Lavi, R., Swamy, C.: Truthful and near-optimal mechanism design via linear programming. In: 46th Annual IEEE Symposium on Foundations of Computer Science (FOCS 2005), pp. 595–604 (2005) 16. Guo, Y., Lim, A., Rodrigues, B., Zhu, Y.: A non-exact approach and experiment studies on the combinatorial auction problem. In: Proc. of HICSS 2005, p. 82.1 (2005)
Deliberation Process in a BDI Model with Bayesian Networks Moser Silva Fagundes1 , Rosa Maria Vicari1 , and Helder Coelho2 1
Instituto de Inform´ atica Universidade Federal do Rio Grande do Sul (UFRGS) Po. B 15.064 91.501-970 Porto Alegre RS Brasil Tel./Fax: +55 51 33166161 {msfagundes,rosa}@inf.ufrgs.br 2 Faculdade de Ciˆencias Universidade de Lisboa Campo Grande 1749-016 Lisboa Portugal [email protected] Abstract. This paper presents a way to perform the deliberation process in a BDI model with Bayesian networks. The assemblage of mental states and Bayesian networks is done by viewing beliefs as networks, and desires and intentions as particular chance variable states that agents pursue. We are particulary concerned with the deliberation about which states of affairs the agents will intend. Perception, planning and execution of plans lie outside the scope of this paper. Our proposal introduces the notion of threshold function to change the agent behavior, and we also discuss the intention selection and the compatibility verification among proactive mental states. Keywords: BDI Architecture, Bayesian Networks, Deliberation.
1
Introduction
Several decisions we have to take in our everyday life demand a large amount of information, which frequently is imprecise and incomplete. Like human beings, resource-bounded cognitive agents that inhabit complex and dynamic environments face the uncertainty phenomena. The AI (Artificial Intelligence) community has made relevant efforts to develop computational models for cognitive agents capable of dealing with uncertainty, since several real world agent applications require that capability. Among these efforts, we include the specification of uncertain beliefs [1], the implementation of intention reconsideration as a partially observable Markov decision process [2], the specification of general probabilistic models of cognition [3,4] and the promotion of interoperability of uncertain knowledge between Bayesian agents [5]. This work is particulary concerned with the deliberation process, responsible for generating the intentions to be pursued by the agent. Efforts have been done in order to enable the deliberation under uncertain conditions. To the best of our knowledge, most researches addressing that issue employ decision theory principles to select intentions (frequently abstracted by the notion of action plans) [6,7]. A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 207–218, 2009. c Springer-Verlag Berlin Heidelberg 2009
208
M.S. Fagundes, R.M. Vicari, and H. Coelho
Our main contribution is the specification of a deliberation process that takes into account probabilistic information through the usage of Bayesian networks to abstract the mental states. Inside our deliberation, the uncertainty is considered to decide if an agent believes that a state can be achieved and if desires and intentions are compatible. The deliberation also uses the causality feature of Bayesian networks to seek the establishment of conditions where desires are feasible. In addition, we discuss the different behaviors (cautious and bold ) that can be obtained by adjusting the parameters used inside the deliberation. The terms cautious and bold agents were previously used by Kinny and Georgeff [8] in their experiments about intention reconsideration and commitment strategies, taking into account the rate of change of the environment. In the context of this work, we use the notions of boldness and cautiousness to qualify the agent’s behavior. The remaining of this paper is organized as follows: section two presents the background for our research; section three presents the assemblage of mental states and Bayesian networks; section four describes the deliberation process; section five presents the related work; and in the section six it is presented our conclusion and future work.
2
Background
We employ Bayesian networks [9,10], also called belief networks, to represent beliefs in our model. They are graphical models that represent causality between variables, allowing the performance of reasoning under conditions of uncertainty in a consistent, efficient, and mathematically sound way. Formally, they consist of directed acyclic graphs and a set of conditional probability distributions. Each node corresponds to a chance variable which has a finite set of mutually exclusive states. The directed arcs specify the causal relation between the variables, and each variable has a conditional probability distribution. In Figure 1, we illustrate the Movies or Soccer Bayesian network, which we use throughout the paper to exemplify how mental states are interpreted by the deliberation process. This very didactic example models the beliefs about the agent going to the movies or playing soccer (beliefs about the agent itself carrying out these activities), and beliefs about the weather and the soccer players (external environment elements). The agent believes that a rainy weather is not a good condition to play soccer, but it is a good call to go to the movies. It also believes that sunny days are good to play soccer, but they were not made to stay in a room watching movies. Finally, our agent believes that soccer games additionally require a group of players. The network and the conditional probability tables illustrated in Figure 1 were specified in order to express these beliefs. Regarding beliefs, it is worth to stress the different interpretations of probabilities in different stages of the Bayesian inference. In Figure 2, we illustrate three stages in the Bayesian inference. The first stage corresponds to the conditional probability table defined by the system designer. The P(Movies|Weather) means the degree of belief (probability) about Movies if a particular state of Weather
Deliberation Process in a BDI Model with Bayesian Networks
209
Fig. 1. Movies or Soccer Bayesian network
Fig. 2. Conditional probability table, computed joint distribution and marginalized table for the variable Movies
occurs with certainty (hard evidence perception). The second stage corresponds to the computed joint distribution for Movies and Weather. The computation was performed taking into account the current probabilities of the states of the variable Weather (P(Weather=rainy)=0.3 and P(Weather=sunny)=0.7). Finally, in the third stage, we obtain the current probability of each state of Movies by marginalizing out Weather of P(Movies,Weather). By this it means that the agent believes with some degree of certainty that a particular state will occur. Throughout the paper, we show how these probabilistic information are used by the deliberation process. Differently from the Bayesian networks, which is a formalized model for representing knowledge, there is not a unique BDI model for agency. In this paper, we specify the deliberation process through algorithms. In order to better contextualize our contribution, we use as starting point the agent control loop specified by Wooldridge [11]. Figure 3 shows the complete control loop, which consists of perception, belief revision function, option selection, filter function, planning, and finally, action plan execution.
210
M.S. Fagundes, R.M. Vicari, and H. Coelho
Fig. 3. BDI Agent control loop [11]
We are particulary concerned with the deliberation, decomposed by the authors in options and filter functions (lines 7–8). The option (desire) selection consists of understanding which desires are available, while the filtering consists of choosing among incompatible desires. For more information about the BDI agent control loop, including more sophisticated versions, see [11].
3
Representing Mental States
Our agents represent the Intentional content of the desires through particular states of variables. Consequently, desires become a subset of beliefs. Differently from beliefs, we use a data structure to make explicit the Intentional content of the desires. By assuming that, we leave outside of the research scope the question ”Where do desires come from?”. In other words, the application designer or domain expert specifies which variable states might be desired. In the Movies or Soccer example, the desires are represented by end nodes, however it is not mandatory. The first requirement for a desire become an intention is the belief support. In other words, the agent will not desire a state unless it believes it is possible to achieve that state. It is done by checking the computed marginal probability (see Figure 2). Consider the example depicted in Figure 4, where the agent will desire the state Movies=yes because it believes the current world conditions (Weather=rainy) support this desire. According to its beliefs, to play soccer is not an option, unless the weather gets sunny. One could argue that our agent will only desire something that is true in the present moment, however this interpretation is not correct. Actually, a high probability associated with a unobserved state does not mean that the state has occurred. It just means that the conditions imposed by the parent variables hold and the state might occur with some degree of certainty. However, in some circumstances, these desired states might never happen unless the agent acts in order to achieve them. A proactive behavior is exhibited by the agent when it pursues states of world where desires become feasible. More specifically, this behavior consists of achieving particular states of parent variables (conditions) based on the values of the conditional probability table of the desired state. They can be viewed as desires connected to a desire through causality.
Deliberation Process in a BDI Model with Bayesian Networks
211
Fig. 4. Supporting desires through beliefs
Fig. 5. Conditions for a desire
In Figure 5, we illustrate an example where Soccer=yes is a desire. It has two conditions, which will be evaluated by the agent (it will decide if it believes or not in the conditions). After evaluating these conditions, the agent realizes that Weather=sunny and Players=yes is the best condition. Assuming that conditions might be seen as desires, it is possible to model conditions of conditions through causal relations, giving rise to a multilevel reasoning. It is important to point out that the desire probability caused by the establishment of the conditions has to be higher than the belief threshold value (minimum probability value required to an agent believe that the state will be observed). For example, it would be a waste of time to bring about the conditions Weather=rain and Player=no because they establish a state of world where the desire Soccer=yes become almost impossible. That threshold issue is discussed in detail in the deliberation process section. Intentions correspond to states of affairs that agents have committed to bring about. This commitment makes the BDI agents to check intentions for incompatibility (mutually exclusive). In our approach, incompatibility is detected when an evidence indicating the achievement of an intention affects negatively other intention by decreasing its chances.
212
M.S. Fagundes, R.M. Vicari, and H. Coelho
Fig. 6. Example of incompatibility among intentions
Figure 6 shows two independent network situations, where an incompatibility is exemplified. In the situation A, the agent believes that the state rainy of Weather has been observed. It increases the chances of Movies=yes and decreases the chances of Soccer=yes. In situation B, the evidence on the state sunny causes the opposite. This incompatibility has resulted from different states that depend on mutually exclusive conditions (in this case, Weather). Even without an evidence on the Weather variable, the achievement of Soccer=yes would decrease the chances of Movies=yes (first by diagnostic reasoning, from Soccer to Weather, and then by causal reasoning from Weather to Movies).
4
Deliberation Process
In the deliberation process BDI agents examine which desires are possible, choose between competing desires, and commit to achieving them. The selection of possible desires done by the options function, illustrated in Figure 7, has two parameters: beliefs and intentions. It begins by obtaining the desires (goals), which are defined in the design-time. After that, the SD (selected desires) variable is initialized. In the line 5, each unintended desire is checked for its compatibility with the current intentions (compatible function call). The next step consists of selecting which desires the agent believes that can be achieved given the current state of world (believe function call, line 6). Desires considered compatible are added to the SD vector (line 7). If a desire d is refused by the believe function, then it is checked by the parent-options function (line 10) in order to establish a state of world where that desire d become feasible. For example, assume that the desire Soccer=yes is refused by the believe function. By checking its conditional probability table (see Figure 1), the agent realizes that the degree of belief about Soccer=yes goes to 0.95 if the conditions Weather=sunny and Players=yes are established. The parent-options function is capable of establishing conditions of conditions by using the causal relations and conditional probability tables of Bayesian networks. We omit here the algorithm of the parent-options function for reasons of space.
Deliberation Process in a BDI Model with Bayesian Networks
213
Fig. 7. Function that selects the options (desires)
Fig. 8. Function to select desires that the agent believes that can be achieved, and function to check desires incompatible with the current intentions
In Figure 8, we show the believe and the compatible functions. The believe function verifies if the computed marginal probability of the desired state is lower than the belief threshold value. If it is lower, then the agent does not believe that state can be achieved. The compatible function has three parameters: the beliefs, the desires to be verified and the intentions. The verification for incompatibility begins by simulating a fake hard evidence indicating the occurrence of the desire. It is done by associating the fake evidence with the node state representing that desire, followed by the performance of a probabilistic inference over the Bayesian networks. The intentions affected by the simulated achievement of the desire are verified with the believe function. If it returns false (the agent does not believe that the intention can be achieved given the current state of world), the desire is considered incompatible. After simulating the achievement of the desire, the
214
M.S. Fagundes, R.M. Vicari, and H. Coelho
Fig. 9. Filter function that selects intentions from competing desires
changes on the network are undone and the beliefs about the current world situation are reestablished. In Figure 9, we show the filter function. It outputs the agent intentions after choosing between competing desires based on beliefs and current intentions. It begins by sorting the desires from lower to higher computer marginal probability (line 39). Here, we employ the same strategy used to detect incompatibility in the compatible function. We simulate the achievement of a desire and verify the remaining ones for incompatibility. Among incompatible desires, we select the one with higher probability according to the sort initially performed. 4.1
Tuning the Agent Behavior
We use the terms cautious and bold to describe the agent behavior in the following three situations inside of the deliberation process: to decide if the agent believes in particular states, to decide if desires are incompatible with intentions and to decide if desires are incompatible among themselves. The first decision that the agent faces in the deliberation process consists of deciding if it believes that a potential desire can be achieved given the current state of the world. This decision, specified in the believe function (Figure 8, line 20), is done by comparing the current value of the computed marginal probability of the potential desire with a threshold value. The threshold indicates the minimum value to believe that the potential desired can be achieved. By increasing the threshold value, we obtain a more cautious behavior since the agent will desire only under high degrees of certainty. By decreasing the threshold value, a bolder behavior is obtained. The last two situations (illustrated in Figure 9 and in Figure 8, respectively) concern with the incompatibility detection. In both cases, the mental states are considered incompatible if the achievement of one affects negatively the
Deliberation Process in a BDI Model with Bayesian Networks
215
achievement of the other by decreasing its chance of success (computed marginal probability). The chance of success is estimated by calling the believe function. But the following question arises: ”How significant is this decrease?”. It depends on the agent’s behavior. A cautious agent will tolerate lower degrees of incompatibility. On the other hand a bold agent will tolerate higher degrees of incompatibility. Outside the deliberation process, the believe function can be used to implement a open-minded agent, since this commitment strategy consists of keeping intentions while the agent believes that it is possible to achieve them. It is worth to point out that the single-minded and the open-minded commitment strategies can be used in the intention selection also. We used the open-minded strategy since the agent has to believe in the state to desire it and consequently to intend it. If we employ a function to verify if the agent believes that the state is impossible, we obtain a single-minded behavior.
5
Related Work
We deal with the integration of two well established areas, namely BDI agency and Bayesian networks. Since the eighties, several significant research concerning these areas have been developed. Among them, we selected the following related ones. A natural question that arises is: ”Why not Influence diagrams [12]?”. To address this question we present the following reasons: – Our approach allows the adjustment of behavior (cautious and bold) in the generation of desires and verification of compatibility among mental states. – The Influence diagrams represent actions within diagrams, however they are limited in the sense that they require the specification of a predefined sequence of actions. It reminds an embedded means-end reasoning, however without advanced planning features. Our approach enables the usage of different planning techniques since we consider means-end reasoning and deliberation as distinct processes. – Bayesian networks are quite simpler than Influence diagrams since the last one include the decision and utility concepts and their edges have different semantics (functional, conditional and informational). From the perspective of our deliberation process, it is required to specify the networks and the potential desires. Concerning uncertainty representation for agency, the related work [5] proposes an architecture for interoperability among Bayesian agents. It specifies an ontology that models the Bayesian network domain, and an agent internal architecture capable of dealing with that ontology. Processes such as Bayesian inference, knowledge manipulation (update and query) and perception handler are proposed. However, this work does not deal with goal-oriented reasoning. In order to allow the implementation of cognitive agents that can deal with uncertain information, the research developed by [1] investigates the usage of
216
M.S. Fagundes, R.M. Vicari, and H. Coelho
Dempster-Shafer theory to incorporate uncertainty in BDI-type agent programming languages. The authors discuss how uncertain beliefs can be represented and reasoned with, and how they can be updated with uncertain information. Despite the fact that this work provides a formal basis to represent uncertain mental states, it does not focus on the deliberation process under uncertainty. The work developed by Rao and Georgeff [6] provides an unifying framework for BDI agents by bringing together theory, systems and applications. The authors transform decision trees, and appropriate deliberation functions, to an equivalent model that represents beliefs, desires and intentions as separate accessibility relations over sets of possible worlds. The resulting model contains two types of information, represented by the probabilities across worlds (beliefaccessibility relation) and the payoffs assigned to paths (desire-accessibility relation). The deliberation function, such as maximizing expected utility, selects one or more best sequences of actions, which correspond to the agent’s intentions. The approach adopted by [6] differentiates from ours in the sense that they employ the expected utility maximization to select actions. We employ uncertain beliefs (variable states and conditional probabilities) to select intentions and check for incompatibilities among mental states. BayesJason [13] is an extension of the Jason [14] tool that addresses the integration of Bayesian networks and BDI agents. Jason is a java-based AgentSpeak(L) [15] interpreter. The BayesJason implements a probabilistic module for Jason, allowing BDI agents to act on uncertain environments. To allow the programming of Bayesian networks within the Jason, the AgentSpeak(L) grammar was modified and extended. Although integrating Bayesian networks and BDI model, the BayesJason uses the Bayesian networks in the context of the intention selection (plans in the AgentSpeak(L) systems). It does not use that knowledge representation to verify the compatibility among mental states, neither to bring about states of world where desires can be achieved.
6
Conclusion and Future Work
We have presented a BDI deliberation process over Bayesian networks, interpreting the mental states through that probabilistic knowledge representation. We focused on the deliberation, leaving outside the research scope issues such as performance measurements, methodology of development, uncertainty on perceptions, and planning. Our proposal of compatibility verification eliminates the necessity of an explicit specification of which mental states are incompatible. By using causal relations and conditional probabilities, the agent infers the degree of incompatibility. Actually, the incompatibilities are implicitly specified in the network, for example in the desires with mutually exclusive conditions. It was highlighted which decisions inside the deliberation can be adjusted to obtain different behaviors under uncertainty, enabling the employment of other Artificial Intelligence techniques, including affective computing [16], by future work to compute the threshold values.
Deliberation Process in a BDI Model with Bayesian Networks
217
Despite the fact that the deliberation does not take into account the meaning of the chance variables, that process is strongly dependent of the conditional probabilities and causal relations. It enables the reuse of that process, but it leaves the application dependent on the designer knowledge and know-how, what happens in other agent development frameworks. We choose to represent intentions separated from plans. By this we mean that our intentions correspond to states of affairs (Intentional contents) instead of sequences of actions that lead the agent to those states. This choice fits exactly in the Wooldridge’s loop, which distinguished the notions of deliberation and means-end reasoning. However, we believe that the serial execution of deliberation and means-end reasoning limits the model in the sense that the agent will have to wait the plan function call to realize that some states can not be achieved. For example, our agent would have to wait for the plan function call to realize that it can not change the weather. Previous work, including PRS-like systems [17], propose plan libraries in order to guarantee the agent reactivity. The same plan libraries could be employed just to check action plans instead of reacting. Future work on the deliberation process improvement is threefold. The first research direction consists of experimenting different AI techniques to implement the threshold function. Currently, we are working on the modification of the Tileworld [18] testbed to experiment our model in that uncertain domain. The second one is concerned with modifications on the control loop, covering commitment strategies, intention reconsideration and scheduling between deliberation and means-end reasoning. The third research direction consists of integrating the Bayesian network ontology [5] in our agent architecture. This integration will make possible to perform an ontological reasoning over the knowledge representation (Bayesian networks) and over the domain knowledge (concepts associated with chance variables, corresponding to the knowledge about the application domain).
References 1. Kwisthout, J., Dastani, M.: Modelling uncertainty in agent programming. In: Baldoni, M., Endriss, U., Omicini, A., Torroni, P. (eds.) DALT 2005. LNCS, vol. 3904, pp. 17–32. Springer, Heidelberg (2006) 2. Schut, M.C., Wooldridge, M., Parsons, S.: Reasoning about intentions in uncertain domains. In: Benferhat, S., Besnard, P. (eds.) ECSQARU 2001. LNCS, vol. 2143, pp. 84–95. Springer, Heidelberg (2001) 3. Tenenbaum, J.B., Griffiths, T.L., Kemp, C.: Theory-based bayesian models of inductive learning and reasoning. Trends in Cognitive Sciences 10(7), 309–318 (2006) 4. Baker, C., Tenenbaum, J.B., Saxe, R.: Bayesian models of human action understanding. In: Neural Information Processing Systems, pp. 99–106 (2005) 5. Santos, E.R., Fagundes, M.S., Vicari, R.M.: An ontology-based approch to interoperability for Bayesian agents. In: Proceedings of the International Conference on Autonomous Agents and Multiagent Systems, Honolulu, Hawaii, pp. 1304–1306. ACM Press, New York (2007)
218
M.S. Fagundes, R.M. Vicari, and H. Coelho
6. Rao, A.S., Georgeff, M.P.: BDI-agents: from theory to practice. In: Proceedings of the 1st International Conference on Multiagent Systems, San Francisco, pp. 312–319 (1995) 7. Ambroszkiewicz, S., Komar, J.: A model of BDI-agent in game-theoretic framework. In: Model Age Workshop, pp. 8–19 (1997) 8. Kinny, D., Georgeff, M.: Commitment and effectiveness of situated agents. In: International Joint Conference on Artificial Intelligence, pp. 82–88 (1991) 9. Pearl, J.: Belief networks revisited. Artificial Intelligence 59(1-2), 49–56 (1993) 10. Jensen, F.V.: Bayesian Networks and Decision Graphs. Springer, Heidelberg (2001) 11. Wooldridge, M.: Reasoning about Rational Agents. Intelligent Robots and Autonomous Agents. MIT Press, Cambridge (2000) 12. Cowell, R.G., Dawid, A.P., Lauritzen, S.L., Spiegelhalter, D.J.: Probabilistic Networks and Expert Systems. Springer, Heidelberg (1999) 13. Calcin, O.G.P.: BayesJason: Bayesian networks and Jason interpreter. In: Iberoamerican Workshop on Multi-Agent Systems (2006) 14. Bordini, R.H., Fisher, M., Visser, W., Wooldridge, M.J.: Verifiable multi-agent programs. In: Dastani, M., Dix, J., El Fallah-Seghrouchni, A. (eds.) PROMAS 2003. LNCS, vol. 3067, pp. 72–89. Springer, Heidelberg (2004) 15. Rao, A.S.: Agentspeak(l): Bdi agents speak out in a logical computable language. In: Perram, J., Van de Velde, W. (eds.) MAAMAW 1996. LNCS, vol. 1038, pp. 42–55. Springer, Heidelberg (1996) 16. Picard, R.A.: Affective Computing. MIT Press, Cambridge (1997) 17. Georgeff, M.P., Ingrand, F.F.: Decision-making in an embedded reasoning system. In: Proceedings of the 11th International Joint Conference on Artificial Intelligence, Detroit, MI, pp. 972–978 (August 1989) 18. Pollack, M.E., Ringuette, M.: Introducing the tileworld: Experimentally evaluating agent architectures. In: Proceedings of the 8th National Conference on Artificial Intelligence, pp. 183–189 (1990)
An Asymmetric Protocol for Argumentation Games in Defeasible Logic Jenny Eriksson Lundstr¨om1, Guido Governatori2, Subhasis Thakur2, and Vineet Padmanabhan3 1
Dept. of Information Science, Computer Science Division, Uppsala University, Sweden 2 School of ITEE, The University of Queensland, Brisbane, Australia 3 Dept. of Computer & Information Sciences, University of Hyderabad, India [email protected], {guido,subhasis}@itee.uq.edu.au, [email protected]
Abstract. Agent interactions where the agents hold conflicting goals could be modelled as adversarial argumentation games. In many real-life situations (e.g., criminal litigation, consumer legislation), due to ethical, moral or other principles governing interaction, the burden of proof, i.e., which party is to lose if the evidence is balanced [22], is a priori fixed to one of the parties. Analogously, when resolving disputes in a heterogeneous agent-system the unequal importance of different agents for carrying out the overall system goal need to be accounted for. In this paper we present an asymmetric protocol for an adversarial argumentation game in Defeasible Logic, suggesting Defeasible Logic as a general representation formalism for argumentation games modelling agent interactions.
1 Introduction Adversarial situations arise when agents in pursuit of their goals interact with other agents pursuing goals of a conflicting nature. In a setting where issues need to be resolved, the agent interaction could be modelled as an adversarial argumentation game. Argumentation games are defeasible, meaning that an argument put forward by one of the agents in support of a conclusion could be defeated by contrary evidence and arguments put forward by the other agent. Thus, the agents are to resolve the dispute by putting forward the arguments that will enable the best outcome for their cases. Using a symmetrical protocol an argumentation game between homogeneous (equally strong) parties could be modelled. However, as in many real-life settings, also in agent systems disputes arise where the claims of the agents involved in the interaction (e.g., regarding distribution of a scarce resource) are of unequal importance to the overall goal of the agent system, and thus need to be handled accordingly. In addition, as in many reallife situations the evidence presented by the parties of a dispute may be inconclusive and the accompanying arguments incoherent. Thus, a majority of the disputes has to be resolved by higher-level principles guiding the interaction. One important principle is referred to as the burden of proof, cf. e.g. [22]. To accommodate for a correct outcome of argumentation games in heterogeneous agent systems, we present an asymmetric protocol for adversarial argumentation games. A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 219–231, 2009. c Springer-Verlag Berlin Heidelberg 2009
220
J. Eriksson Lundstr¨om et al.
The paper is organized as follows: In Section 2 we present argumentation games and their setup. In Section 3 we highlight the most relevant features of defeasible logic. We discuss the formalization of argumentation games using defeasible logic as presented in [25] in Section 4. Section 5 presents the asymmetrical protocol for argumentation games. We use a criminal litigation setting to illustrate and discuss some of the benefits of the model. Section 6 presents some related work. In Section 7 we conclude.
2 Argumentation Games Consider an adversarial argumentation (dialogue) game as an interaction between two parties, the Proponent and the Opponent. The two parties debate over a topic. Each equipped with a set of arguments the parties take turn in putting forward a subset of these arguments, i.e., move, with the sole purpose of justifying their claim. The game is governed by a protocol for admissible moves and the winning conditions. For the proponent a basic protocol for an argumentation game is that the arguments of the move attack the previous move of the adversary, and that the main claim follows from the arguments assessed as currently valid. For the opponent goes that an admissible move has to attack the previous move of the adversary, and that the main claim is not derivable. Even though more complex winning criteria could be devised, by a basic protocol, a player wins the argumentation game when the other party is out of admissible moves.
3 Defeasible Logic Defeasible Logic (DL) [18,1,2] is a simple, flexible, rule based non-monotonic formalism able to capture different and sometimes incompatible facets of non-monotonic reasoning [3], and efficient and powerful implementations have been proposed [17,12]. Knowledge in DL can be represented in two ways: facts and rules. Facts are represented either in form of states of affairs (literal and modal literal) and actions that have been performed. Facts are represented by predicates. For example, “Tweety is a penguin” is represented by Penguin(Tweety). A rule describes the relationship between a set of literals (premises) and a literal (conclusion), and we can specify how strong the relationship is. As usual rules allow us to derive new conclusions given a set of premises. We distinguish between strict rules, defeasible rules and defeaters represented, respectively, by expressions of the form A1 , . . . , An → B, A1 , . . . , An ⇒ B and A1 , . . . , An ; B, where A1 , . . . , An is a possibly empty set of prerequisites (causes) and B is the conclusion (effect) of the rule. We only consider rules that are essentially propositional. Rules containing free variables are interpreted as the set of their ground instances. Strict rules are rules in the classical sense: whenever the premises are indisputable then so is the conclusion. Thus they can be used for definitional clauses. An example of a strict rule is “Penguins are birds”, formally: Penguin(X) → Bird(X). Defeasible rules are rules that can be defeated by contrary evidence. An example of such a rule is “Birds usually fly”: Bird(X) ⇒ Fly(X). The idea is that if we know that X is a bird, we may conclude that X can fly unless other evidence suggest she may not.
An Asymmetric Protocol for Argumentation Games in Defeasible Logic
221
Defeaters are a special kind of rules. They are used to prevent conclusions, not to support them. For example: Heavy(X) ; ¬Fly(X). This rule states that if something is heavy then it might not fly. This rule can prevent the derivation of a “fly” conclusion. On the other hand it cannot be used to support a “not fly” conclusion. DL is a “skeptical” non-monotonic logic, meaning that it does not support contradictory conclusions. Instead DL seeks to resolve conflicts. In cases where there is some support for concluding A but also support for concluding ¬A, DL does not conclude either of them (thus the name “skeptical”). If the support for A has priority over the support for ¬A then A is concluded. No conclusion can be drawn from conflicting rules in DL unless these rules are prioritized. The superiority relation among rules is used to define priorities among rules, that is, where one rule may override the conclusion of another rule. For example, given the defeasible rules r : Bird(X) ⇒ Fly(X)
r : Penguin(X) ⇒ ¬Fly(X)
which contradict one another, no conclusive decision can be made about whether a Tweety can fly or not. But if we introduce a superiority relation with r r, then we can indeed conclude that Tweety cannot fly since it is a penguin. We now give a short informal presentation of how conclusions are drawn in DL. Let D be a theory in DL (i.e., a collection of facts, rules and a superiority relation). A conclusion of D is a tagged literal and can have one of the following four forms: +Δ q −Δ q +∂ q −∂ q
meaning that q is definitely provable in D (i.e., using only facts and strict rules). meaning that we have proved that q is not definitely provable in D. meaning that q is defeasibly provable in D. meaning that we have proved that q is not defeasibly provable in D.
Strict derivations are obtained by forward chaining of strict rules, while a defeasible conclusion p can be derived if there is a rule whose conclusion is p, whose prerequisites (antecedent) have either already been proven or given in the case at hand (i.e., facts), and any stronger rule whose conclusion is ¬p has prerequisites that fail to be derived. Formally a DL theory (as formalized by [5]) is a structure D = (F, R, ) where F is a finite set of factual premises, R a finite set of rules, and a superiority relation on R. Given a set R of rules, we denote the set of all strict rules in R by Rs , the set of strict and defeasible rules in R by Rsd , the set of defeasible rules in R by Rd , and the set of defeaters in R by Rd f t . R[q] denotes the set of rules in R with consequent q. In the following ∼p denotes the complement of p, that is, ∼p is ¬q if p = q, and ∼p is q if p is ¬q. For a rule r we will use A(r) to indicate the body or antecedent of the rule and C(r) for the head or consequent of the rule. A rule r consists of its antecedent A(r) (written on the left; A(r) may be omitted if it is the empty set) which is a finite set of literals, an arrow, and its consequent C(r) which is a literal. Provability is based on the concept of a derivation (or proof) in D. A derivation is a finite sequence P = (P(1), . . . , P(n)) of tagged literals. Each tagged literal satisfies some proof conditions. A proof condition corresponds to the inference rules corresponding to one of the four kinds of conclusions we have mentioned above. P(1..i) denotes the
222
J. Eriksson Lundstr¨om et al.
initial part of the sequence P of length i. Here we state the conditions for strictly and defeasibly derivable conclusions (see [1] for the full presentation of the logic): If P(i + 1) = +Δ q then (1) q ∈ F, or (2) r ∈ Rs [q], ∀a ∈ A(r) : +Δ a ∈ P(1..i). If P(i + 1) = +∂ q then (1) +Δ q ∈ P(1..i), or (2) (2.1) ∃r ∈ Rsd [q]∀a ∈ A(r) : +∂ a ∈ P(1..i) and (2.2) −Δ ∼q ∈ P(1..i) and (2.3) ∀s ∈ R[∼q] either (2.3.1) ∃a ∈ A(s) : −∂ a ∈ P(1..i) or (2.3.2) ∃t ∈ Rsd [q] ∀a ∈ A(t) : +∂ a ∈ P(1..i) and t s.
4 Dialogue Games in Defeasible Logic – A Symmetric Protocol In [25] we presented a model for an argumentation game in DL using a basic symmetric protocol for adversarial dispute. We parse a dialogue into defeasible rules utilizing time of the dialogue as the time of the rule. In order to resolve the dispute, the agents take turn in putting forward arguments from a private knowledge base, i.e., a finite set of (defeasible) arguments in support of their claim. At each time step, an agent is allowed to put forward any of its arguments (rules) that has precedence over any contradictory defeasible rule of the previous steps. In this symmetric protocol we assume that if at time t2 we have a valid rule w ∈ Rtsd2 which contradicts a defeasible rule s ∈ Rtd1 of time t1 and t2 > t1 then the strength of w is greater than s. The expression a@t denotes that the expression a being put forward or upgraded at time t.
(w s)@t iff (w, s) ∈ or w ∈ Rt [P], s ∈ Rt [¬P] where t < t A common public knowledge base holds the common knowledge, which is a theory in defeasible logic. The sets of agreed common knowledge construct the theories T1 , T2 , . . . , Tn respectively as the undefeated defeasible rules from the previously adjacent step are strengthened into strict rules and the defeated defeasible rules are removed. Thus, if P is the conclusion of a defeasible rule of the adjacent previous step ti−1 , regardless of its origin, the agreed common knowledge is created by strengthening the status of rules from defeasible to strict in the adjacent next step ti+1 if:
∃r ∈ Rtd [P] t < t ∀t : t < t < t Rtsd [¬P] = 0/ and ∀a ∈ A(r) : +Δ a@t The proof procedures of the defeasible logic are applied to the critical literal at each time step, thus determining the burden of proof and the outcome of the argumentation game. The first theory T1 = ({}, R1d , ) is created from the arguments (ARG1 ) presented by the first player, and the second theory T2 = ({}, R2d , ) is created through modifications of T1 by the arguments (ARG2 ) presented by player 2. The transition rules from the first theory to the second theory were devised as follows:
An Asymmetric Protocol for Argumentation Games in Defeasible Logic
223
1. If r ∈ R1d and ∀s ∈ ARG2 , ¬C(s) = C(r) ∧ ¬C(s) ∈ / A(r), then r ∈ R2d . 2. All rules of ARG2 are added to T2 as defeasible rule. Here we assume that ARG2 is valid and that a valid argument, by the above defined precedence relations, is stronger than any contradictory argument of the previous step. At time n, n > 2 theory Tn is created through modification of Tn−1 by arguments (ARGn ) of the player who has to play at that step. The rules for transition from Tn−1 to Tn are 1. If r ∈ Rn−1 then r ∈ Rns . s and ∀s ∈ ARGn−1 , ¬C(s) = C(r) ∧ ¬C(s) ∈ / A(r), then r ∈ Rns ; otherwise 2. If r ∈ Rn−2 d r∈ / Rn . Here we should note that the player will not oppose its previous argument. Thus, all unchallenged rules of time n − 2 are upgraded as strict rules at time n. 3. If r ∈ Rn−1 and ∀s ∈ ARGn−1 , ¬C(s) = C(r) ∧ ¬C(s) ∈ / A(r), then r ∈ Rnd . Unchald lenged defeasible rules of time n − 1 are added to Tn as defeasible rules. 4. All rules of ARGn−1 are added to Tn as defeasible rules. Here we assume that ARGn−1 is valid and that a valid argument, by the above defined precedence relations, is stronger than any contradictory argument of previous step. The winning criteria for a basic game are devised as an agent to be winning if the claim q is definitely proven +Δ q at any time step. If an agent at any step of the game proves +∂ A the burden of production as well as persuasion of −∂ A, or −Δ A or +Δ ¬A or +∂ ¬A are placed on the other party. Using a symmetric protocol, a dispute between equally strong parties could be modelled as an argumentation game and resolved accordingly. However, in many situations ethical, moral or other reasons (cf., e.g., criminal litigation1 ) advocate for special concerns to be taken on behalf of one of the parties. To accommodate such settings, asymmetric protocols are required.
5 Dialogue Games in Defeasible Logic – An Asymmetric Protocol Here we present an asymmetric model for adversarial argumentation games between two parties: the Prosecutor and the Defendant. As in the symmetric protocol, we parse the dialogue into defeasible rules utilizing time of the dialogue as the time of the rule. Each agent has at its disposition a private knowledge base consisting of a finite set of defeasible arguments in support of their claim (the critical literal). Initiated by the prosecutor, the parties take turns in presenting their arguments. At each time step the proof procedures are applied to the critical literal. The outcome of an argumentation game is determined by the final stage of the game, whilst the intermediate stages are illustrating the situation for the current situation. For common sense reasons, as an argument put forward cannot be revoked from impacting the argumentation, we do not allow for backtracking. The winning criteria for a basic game are devised as an agent to be winning if the claim q is definitely proven +Δ q at any time step. However, analogously to the burden 1
“Homo praesumitur bonus donec probetur malus” lat: Innocent until proven guilty. The adoption of this presumption of innocence in many national statutes results in that the defendant of a criminal litigation only is required to at most produce an exception to the accusation.
224
J. Eriksson Lundstr¨om et al.
of persuasion, which imposes a requirement of providing a justified (i.e., strongly defeating) argument for the issue on which the burden rests (based on rebutting defeat) [22], we require of the prosecutor a strong defeat of any argument (including the critical literal) presented by the defendant. Thus, if the prosecutor at any step of the game proves +∂ A, the prosecutor still holds the burden to produce proof of +Δ A in order to win. For the defendant only a burden of production of an exception +∂ ¬A, (being subsumed by −∂ A, −Δ A or +Δ ¬A) is imposed. If the defendant at any step of the game proves the exception +∂ ¬A, the burden of persuasion placed on the prosecutor necessitates the proof of +Δ A (including −∂ ¬A) in order for the prosecutor to win. In the symmetric protocol, regardless of its origin, time brings strengthening of undefeated rules from defeasible to strict. Here we require that the strengthening of rules originating from the prosecutor only occurs when the rule could be derived from arguments already put forward by the defendant. In other cases, undefeated rules from the previously adjacent step presented by the prosecutor remain as defeasible rules in the common knowledge base. Defeated rules are removed at each step. As we do not allow the prosecutor to repeat arguments and the arguments put forward have to strongly defeat any arguments put forward by the defendant, the game will terminate. In the symmetric protocol at each step any agent in turn to move can present an argument if its strength is stronger than contradictory defeasible rules of the previous steps. In our asymmetrical distribution of the burden of proof, the defendant is allowed to present an argument that merely weakly defeats the argument of the prosecutor of the previous steps. As consequence the defendant could remain with the same argument for fulfilling his burden of production of an assumption +∂ ¬A as response to +∂ A. The strength of an argument is determined by either previously known superiority relationships or validity of that rule. Adhering to the above presented syntax, we write t yi ∈ Rxj |x ∈ {d,s,sd}. Here y is a rule identifier with the subscripts i ∈ {p, d} where p means that the origin of the rule is the prosecutor and d means that the origin of the rule is the defendant. In the following, unless needed, the indexing is left out for readability reasons. If ∀r ∈ Rtsi [q] and ∀s ∈ Rtdi [¬q], then (r s)@ti We consider that the rule strength of a strict rule is greater than the rule strength of a defeasible rule.
(wd s p )@t iff (wd , s p ) ∈ or s p ∈ Rt [P] and wd ∈ Rt [¬P], where t < t < t For defeasible rules presented by the defendant we simply assume that if at time t2 we have a valid rule wd ∈ Rt2 which contradicts a defeasible rule s p ∈ Rt1 of time t1 and t2 > t1 then the strength of wd is greater than s p . This fits well with the burden of persuasion being placed on the prosecutor. We utilize defeasible logic to determine strength of a new rule presented by the players.
(s p wd )@t iff (s p , wd ) ∈ or wd ∈ Rt [¬P] and s p ∈ Rt [P] and ∀a ∈ A(s) : +Δ a@t, where t < t < t else (wd s p )@t iff (wd , s p ) ∈ or wd ∈ Rt [¬P] and s p ∈ Rt [P] and ∃a ∈ A(s) : −Δ a@t, where t < t < t
An Asymmetric Protocol for Argumentation Games in Defeasible Logic
225
As the prosecutor holds the burden of persuasion, we assume that unless the rule priority is set, that only if at time t2 we have a valid rule s p ∈ Rt2 which contradicts a defeasible rule wd ∈ Rt1 of time t1 and t2 > t1 and the rule presented by the prosecutor strongly defeats the rule of the defendant then the strength of the argument s p of the prosecutor is greater than the argument of the defendant wd . In all other situations the opposite goes, thus rendering the strength of the argument s p of the prosecutor weaker than the argument of the defendant wd . In this asymmetric protocol the criteria for strengthening the rule strength of a defeasible rule to a strict rule are devised as follows:
∃r ∈ Rtd [P] t < t ∀t : t < t < t Rtsd [¬P] = 0/ and ∀a ∈ A(r) : +Δ a@t
If P is the conclusion of a defeasible rule r ∈ Rtd of the adjacent previous step t and the rule was presented by the defendant then we can upgrade the rule status from defeasible to strict in the next time step t if no counterarguments are presented by the prosecutor at time t . For t Even < t Odd < t Odd
Even
∃r ∈ Rtd [P], Rtsd [¬P] = 0/ and ∀a ∈ A(r) : +Δ a@t and Even 1) ∃r ∈ Rtds [P] and ∀a ∈ A(r) : +Δ a@t, or Even 2) +∂ [P]@t f romRtds Odd
However, if P is the conclusion of a defeasible rule r ∈ Rtd of the adjacent previous step t Odd and the rule was presented by the prosecutor then we can upgrade the rule status from defeasible to strict in next step only in the case of no counterarguments being presented by the defendant at the adjacently following time t Even and the defeasible (or strict) rule r has been put forward by the defendant or the conclusion P follows Even defeasibly from the defeasible or strict rules Rtds presented by the defendant at the adjacently following time t Even . An argumentation game is initiated at time 1 by the prosecutor putting forward arguments (ARG1 ) from its private knowledge into the common knowledge base to prove its claim (critical literal) A. As the parties take turns in presenting their arguments, at time 2 the defendant agent responds to the accusations. We allow arguments in the form of valid defeasible rules being as strong or stronger than at least some rules of theory T1 . The common sets of argument construct the theories T1 , T2 , T3 , . . . , Tn respectively where the subscripts indicate the time at which the common sets of argument are constructed. As all arguments in the private knowledge base of the agents are defeasible, in the first two theories the common set of arguments consists only of defeasible rules from both time 1 and time 2, according to the following transition rules: Let the first theory T1 = ({}, R1d , ) be created from arguments (ARG1 ), the operative plea of prosecutor, and the second theory T2 = ({}, R2d , ) be created through modifications of T1 by arguments (ARG2) from the defendant. Now the transition rules from the first theory T1 to the second theory T2 are as follows: / A(r), then r ∈ R2d . 1. If r ∈ R1d and ∀s ∈ ARG2 , ¬C(s) = C(r) ∧¬C(s) ∈ 2. All rules of (ARG2 ) are added to T2 as defeasible rules. Under the assumption of (ARG2 ) be valid and that, by the above defined precedence relations, any valid argument from the defendant is stronger than its contradictory argument (from the
226
J. Eriksson Lundstr¨om et al.
prosecutor) of the adjacent previous step. As all unchallenged rules of the prosecutor are added to T2 as defeasible rules T2 now consists of all unchallenged rules of the prosecutor and all arguments (ARG2 ) of the defendant. At time n = 2m + 1 (m > 0), theory Tn is created through modification of Tn−1 by arguments (ARGn ) of the prosecutor. Accounting for the heterogeneity of the parties we capture the asymmetrical burden of proof by the following rules for transition from theories Tn−1 to Tn : 1. If r ∈ R1d and ∀s ∈ ARGn−1 , ¬C(s) = C(r) ∧ ¬C(s) ∈ / A(r), and 1) r ∈ Rn−1 or 2) d n−1 n Rd C(r), then r ∈ Rs . and ∀s ∈ ARGn−1, ¬C(s) = C(r) ∧ ¬C(s) ∈ / A(r), then r ∈ Rnd . Here we 2. If r ∈ Rn−2 d should note that, in contrast to the symmetric protocol, even though the defendant has not actively challenged these arguments and the prosecutor will not oppose its previous argument by the rules of the game, we find it to be a too strong presumption to strengthen the rule status of these rules to strict rules. Thus, unless the argument is acknowledged by the defendant (see transition rule 1.), all unchallenged rules of time n − 2 of the prosecutor remain as defeasible rules at time n. and ∀s ∈ ARGn , ¬C(s) = C(r) ∧ ¬C(s) ∈ / A(r), then r ∈ Rnd . All unchal3. If r ∈ Rn−1 d lenged defeasible rules of time n − 1 (originating from the defendant) are added as defeasible rules at time n. 4. If r ∈ Rn−1 and ∀s ∈ ARGn , ¬C(s) = C(r) ∧ ¬C(s) ∈ / A(r), and r s then r ∈ R3d . d For removal is required that all rules of the defendant have to be strongly defeated by the prosecutor. Thus, also the defeasible rules of time n − 1 (originating from the defendant) of equal or stronger strength are added as defeasible rules at time n. / A(r), then r ∈ Rnd . All unchal5. If r ∈ ARGn and ∀s ∈ Rn−1 d , ¬C(s) = C(r) ∧ ¬C(s) ∈ lenged rules of (ARGn ) are added to Tn as defeasible rules. 6. If r ∈ ARGn and ∀s ∈ Rn−1 / A(r) and r s, then r ∈ Rnd . All d , ¬C(s) = C(r)∧¬C(s) ∈ rules of (ARGn ) that are of higher priority, i.e., strongly defeat the arguments of the defendant are added to Tn as defeasible rules. Here due to the burden of production of the prosecutor, all arguments added are required to either be unchallenged or to strongly defeat all previous arguments of the defendant. This way, by putting forward new arguments, the prosecutor could strengthen its claim. As a result Tn consists of the unchallenged defeasible rules of Tn−2 of the prosecutor, the unchallenged defeasible rules and the rules of by Tn−1 of the defendant that are challenged by (ARGn ) but found equally strong or stronger, and the unchallenged defeasible rules of the prosecutor from (ARGn ). At time n = 2m (m > 1), theory Tn is created through modification of Tn−1 by arguments (ARGn ) of the defendant. The transitions from Tn−1 to Tn are devised as follows: 1. If r ∈ Rn−1 and ∀s ∈ ARGn , ¬C(s) = C(r) ∧ ¬C(s) ∈ / A(r), then r ∈ Rns . The defend dant will not oppose its previous argument by the rules of the game. Thus, all the unchallenged defeasible rules Rn−1 are upgraded as strict rules, i.e., facts, at time d n. and ∀s ∈ ARGn , ¬C(s) = C(r) ∧ ¬C(s) ∈ / A(r), then r ∈ Rnd . All unchal2. If r ∈ Rn−1 d lenged defeasible rules Rn−1 are added as defeasible rules at time n. As already d stated, the defendant cannot challenge her own rules presented in Rn−1 d .
An Asymmetric Protocol for Argumentation Games in Defeasible Logic
227
3. All rules of (ARGn ) are added to Tn as defeasible rules. In contrast to arguments, e.g., (ARGn−1 ) originating from the prosecutor, as (ARGn ) originates from the defendant we merely require that all rules of (ARGn) are valid and at least as strong (or stronger) than any of its contradictory arguments presented by the prosecutor. An Example – Presumption of Innocence. The asymmetric model of argumentation game defeasible logic is illustrated by elaboration on the example of [25]. Consider a particular argumentation game between the prosecutor Alice and the defendant Bob. Alice is trying to convict Bob by proving A and Bob is claiming ¬A. At each step they maintain a current set of rules (CRt , where t is the time). Here a rule consists of its name R i (where R i indicates that the rule belongs to the current set CRt as opposed to the rules present in the private knowledge bases of the parties denoted by Ri ), its antecedent A(r), which is a finite set of literals, an arrow, its consequent C(r), which is a literal and @tx |x ∈ {1, 2, 3, . . ., n} denotes the time of the rule, which is updated at each step. Alice initiates the game at time t1 by presenting her first move as R1 : 0/ ⇒ B
R2 : B ⇒ A
This will generate two defeasible rules as R 1(0/ ⇒ B)@t1 , R 2(B ⇒ A)@t1 . Thus at time t1 CRt1 =[R 1, R 2] and we can prove +∂ A@t1 . Now at next time point, Bob gets its chance to disprove A. At time t2 , Bob presents the following argument, R3 : 0/ ⇒ D
R4 : D ⇒ ¬A
This will generate two new defeasible rules as R 3(0/ ⇒ D)@t2 R 4(D ⇒ ¬A)@t2 . Now, Bob only attacks R 2 presented by Alice at previous step by R 4 and R 2 is removed from CRt2 . Note that as t2 > t1 , the strength of R 4 is greater than R 2 according to the strength determination rule for the defendant. At time t2 , R 1 remains unchallenged but as this rule is not utilized even as a premise in the reasoning of Bob it does not commit Bob to this rule, (leaving Bob the possibility to dismiss this rule by contesting it at a later time or Alice to present evidence to strengthen this rule by the strength determination rule). Thus, the rule remain in CRt2 as R 1(0/ ⇒ B)@t2 . This is in contrast to the symmetric protocol where R 1 as unchallenged is changed to a strict rule R 1(0/ → B)@t2 (a fact) regardless of its origin. Note that we change the time stamp of the rule from t1 to t2 to indicate that it is member of CRt2 at time t2 . Thus, CRt2 =[R 1, R 3, R 4]. The proof at time t2 is +∂ ¬A (which implies that we also have −Δ A as the latter rule R 4 is stronger than R 2 in accordance to the first strength determination rule. Next at time t3 , in order to defeat the arguments presented by Bob, Alice presents the following arguments: R5 : B ⇒ ¬D
R6 : 0/ ⇒ E
R7 : E ⇒ A
So the translated defeasible rules are R 5(B ⇒ ¬D)@t3 , R 6(0/ ⇒ E)@t3 , R 7(E ⇒ A)@t3 . Now the CRt3 is [R 1, R 3, R 4, R 6] as the rule R 4 is stronger than R 7 and the rule R 3 is stronger than R 5 according to the transition rule as neither the argument R 7 nor R 5 can strongly defeat R 4 or R 3 respectively and thus they are removed. So the proof at this time point remain +∂ ¬A. If Alice cannot present any additional arguments strongly defeating R 4 in the next step the rule R 4 is strengthened into a
228
J. Eriksson Lundstr¨om et al.
strict rule resulting in the proof of +Δ ¬A. Thus, in contrast to the symmetric protocol example where Alice wins the game as she were able to upgrade the defeasible rules supporting the proof of +∂ A by use of the rule priority assumption in 5, she would need the rule to be strongly defeated as e.g by addition of a rule R i : E ⇒ ¬D to prevent the rule R 4 from being strengthened into a strict rule at time t4 . As this is not the case, Bob wins the game at time t4 and is acquitted from the criminal charge of A. Another Example – Beyond Reasonable Doubt. Consider a second argumentation game between the prosecutor Alice and the defendant Bob. Alice is still trying to convict Bob by proving A and Bob is claiming ¬A. Again Alice initiates the game at time t1 by presenting her first move as R1 : 0/ ⇒ B
R2 : F ⇒ A
This will generate two defeasible rules as R 1(0/ ⇒ B)@t1 , R 2(F ⇒ A)@t1 . Thus at time t1 CRt1 =[R 1, R 2] and we have proof +∂ A@t1 . Now at next time point, Bob gets its chance to disprove A. At time t2 (t2 > t1 ), Bob presents the following argument, R3 : E ⇒ D
R4 : B, D ⇒ ¬A
This will generate two new defeasible rules as R 3(E ⇒ D)@t2 , R 4(B, D ⇒ ¬A)@t2 . Now, Bob only attacks R 2 presented by Alice at previous step by R 4 and R 2 is removed from CR. As t2 > t1 , the strength of R 4 is greater than R 2 according to the strength determination rule for the defendant. Thus, CRt2 =[R 1, R 3, R 4]. The proof at time t2 is −∂ A, which includes that we also have −Δ A as the latter rule R 4 is stronger than R 2. Next at time t3 , Alice presents the following arguments, R5 : ¬E ⇒ ¬D
R6 : 0/ ⇒ ¬E
R7 : B ⇒ A
The translated defeasible rules are R 5(¬E ⇒ ¬D)@t3 , R 6(0/ ⇒ ¬E)@t3 , R 7(B ⇒ A)@t3 . Now CRt3 =[R 1, R 4, R 5, R 6, R 7] as R 3 is strongly defeated by R 5 and R 6 and thus, it is removed. At time t3 , R 1 remains unchallenged and as it is utilized as a premise in the reasoning of Bob and thus commits Bob to this rule (which is justified as Bob could not be allowed to rely on not actively presented inconsistencies), presented by the prosecutor Alice it is strengthened into a strict rule (i.e., a fact) as R 1(0/ → B)@t3 . So the proof at this time point is +∂ A as the rule is stronger according to the transition rules. If Bob does not present valid arguments in the next step Alice wins the game as from Bob’s argumentation her claim is corroborated.
6 Related Work In this paper we augment the model of dialogue games in defeasible logic of [25]. The work is based on ALIS [22], and we are inspired by [14] as we have separated the knowledge of the players into (1) private knowledge and (2) public knowledge. The common public knowledge forms the common set of arguments, which is a theory in
An Asymmetric Protocol for Argumentation Games in Defeasible Logic
229
defeasible logic. As such our model in contrast to [22] provides a closer approximation of argumentation games for the agent setting, as the agent could choose at what time and which parts of its arguments (i.e., private knowledge) be disclosed to its opponent. Substantial work have been done on argumentation games in the AI and Law-field. [4] presents an early specification and implementation of an argumentation game based on the Toulmin argument-schema without a specified underlying logic. [8] presented The Pleadings Game as a normative formalization and fully implemented computational model, using conditional entailment. The goal of the model was to identify issues in the argumentation rather than as in our case elaborating on the status of the main claim. The dialectic proof procedures presented by [7] focus on minimizing the culprit of argumentation. The proof procedures are expressed as metalogic programs. Our use of defeasible logic establishes a difference as to the syntactic limitations as the approach in [7] is built on assumptions that are atomic, whereas in our framework the arguments are expressible as rules of propositional defeasible logic, not being directly applicable in the abstract argumentation framework underlying [6]. DiaLaw [16] is a two player game, in which both players make argument moves. The model combines exchange of statements and exchange of arguments, dealing with rhetorical as well as psychological issues of argumentation. However, the main focus for the two players is to convince each other rather than defeating an adversarial as in our case. The abstract argumentation systems of [26,27] study arguments as the object of defeat. The results however are more related to stable semantics than sceptical as in the defeasible logic utilized in our framework and the study is devised as meta games for changing the rules of argumentation games. Modelling argumentation games in defeasible logic has been addressed by [14,13,15]. [15], in contrast to our work, focuses on persuasion dialogues in a cooperative setting. It includes in the process cognitive states of agents such as knowledge and beliefs, and presents some protocols for some types of dialogues (e.g. information seeking, explanation, persuasion). The main reasoning mechanism is based on basic defeasible logic (cf. Section 3), while ignoring recent development in extensions of defeasible logic with modal and epistemic operators for representing the cognitive states of agents [10,11]. [14] provides an extension of defeasible logic to include the step of the dialogue in a way very similar to what we have presented in the paper. A main difference is that the resulting mechanism just defines a metaprogram for an alternative computational algorithm for ambiguity propagating defeasible logic while the logic presented here is ambiguity blocking. In [13], the authors focus on rule sceptic characterizations of arguments and propose the use of sequences of defeasible (meta) theories, while using meta-reasoning (meta-rules or high level rules) to assess the strength of rules for the theories at lower levels.
7 Conclusion In this paper Defeasible Logic is used to capture an asymmetric protocol for argumentation games. We have shown that our model provides for a closer approximation of argumentation games for heterogeneous agent settings. The agent characteristics or the agents’ relative importance in fulfilling the overall goal of the system could be captured,
230
J. Eriksson Lundstr¨om et al.
while all the same the agent is allowed to argue its case in the best way it knows, e.g. choosing at what time any subset of its arguments (i.e. private knowledge) be disclosed to its adversary.
References 1. Antoniou, G., Billington, D., Governatori, G., Maher, M.J.: Representation results for defeasible logic. ACM Transactions on Computational Logic 2(2), 255–287 (2001) 2. Antoniou, G., Billington, D., Governatori, G., Maher, M.J.: Embedding Defeasible Logic in Logic Programming. Theory and Practice of Logic Programming 6(6), 703–735 (2006) 3. Antoniou, G., Billington, D., Governatori, G., Maher, M.J., Rock, A.: A family of defeasible reasoning logics and its implementation. In: Proc. ECAI 2000, pp. 459–463 (2000) 4. Bench-Capon, T.J.M.: Specification and Implementation of Toulmin Dialogue Game. In: Proc. JURIX 1998, Nijmegen, GNI, pp. 5–20 (1984) 5. Billington, D.: Defeasible logic is stable. J. Log. and Comput. 3, 370–400 (1993) 6. Dung, P.M.: On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic programming, and n–person games. Artif. Intel. 77, 321–357 (1995) 7. Dung, P.M., Kowalski, R.A., Toni, F.: Dialectic proof procedures for assumption-based, admissible argumentation. Artif. Intel. 170, 114–159 (2006) 8. Gordon, T.: The Pleadings Game: An artificial Intelligence Model of Procedural Justice. Artif. Intel. and Law 2(4) (1993) 9. Governatori, G., Maher, M.J., Billington, D., Antoniou, G.: Argumentation semantics for defeasible logics. J. Log. and Comput. 14(5), 675–702 (2004) 10. Governatori, G., Rotolo, A.: Defeasible logic: Agency, intention and obligation. In: Lomuscio, A., Nute, D. (eds.) DEON 2004. LNCS (LNAI), vol. 3065, pp. 114–128. Springer, Heidelberg (2004) 11. Governatori, G., Rotolo, A., Padmanabhan, V.: The cost of social agents. In: Proc. AAMAS 2006, pp. 513–520. ACM Press, New York (2006) 12. Grosof, B.N.: Representing e-commerce rules via situated courteous logic programs in RuleML. Electronic Commerce Research and Applications 3(1), 2–20 (2004) 13. Eriksson Lundstr¨om, J., Hamfelt, A., Fischer Nilsson, J.: A Rule-Sceptic Characterization of Acceptable Legal Arguments. In: Proc. ICAIL 2007, pp. 283–284. ACM Press, New York (2007) 14. Hamfelt, A., Eriksson Lundstr¨om, J., Fischer Nilsson, J.: A metalogic formalization of legal argumentation as game trees with defeasible reasoning. In: Proc. ICAIL 2005, pp. 250–251. ACM Press, New York (2005) 15. Letia, I.A., Varic, R.: Defeasible protocols in persuasion dialogues. In: Proc. WE-IAT 2006. IEEE, Los Alamitos (2006) 16. Lodder, A.R.: DiaLaw: On Legal Justification and Dialogical Models of Argumentation. Artif. Intel. and Law 8(2-3), 265–276 (2000) 17. Maher, M.J., Rock, A., Antoniou, G., Billington, D., Miller, T.: Efficient defeasible reasoning systems. International Journal of Artificial Intelligence Tools 10(4), 483–501 (2001) 18. Nute, D.: Defeasible logic. In: Handbook of Logic in Artificial Intelligence and Logic Programming, vol. 3, pp. 353–395. Oxford University Press, Oxford (1994) 19. Prakken, H.: Modelling defeasibility in law:logic or procedure? Fundamenta informaticae, 253–271 (2001) 20. Prakken, H., Sartor, G.: Rules about rules: Assessing conflicting arguments in legal reasoning. Artif. Intel. and Law, 331–368 (1996)
An Asymmetric Protocol for Argumentation Games in Defeasible Logic
231
21. Prakken, H., Sartor, G.: Presumptions and burdens of proof. In: Proc. Jurix 2006, pp. 21–30. IOS Press, Amsterdam (2006) 22. Prakken, H., Sartor, G.: Formalising arguments about the burden of persuation. In: Proc. ICAIL 2007, pp. 97–106. ACM Press, New York (2007) 23. Prakken, H.: Relating protocols for dynamic dispute with logics for defeasible argumentation. Synthese 127, 187–219 (2001) 24. Roth, B., Riveret, R., Rotolo, A., Governatori, G.: Strategic Argumentation: A Game Theoretical Investigation. In: Proc. ICAIL 2007, pp. 81–90. ACM Press, New York (2007) 25. Thakur, S., Governatori, G., Padmanabhan Nair, V., Eriksson Lundstr¨om, J.: Dialogue games in defeasible logic. In: Orgun, M.A., Thornton, J. (eds.) AI 2007. LNCS (LNAI), vol. 4830, pp. 497–506. Springer, Heidelberg (2007) 26. Vreeswijk, G.: Abstract Argumentation Systems. Artif. Intel. 90, 225–279 (1997) 27. Vreeswijk, G.: Representation of formal dispute with a standing order. Artif. Intel. and Law 8(2/3), 205–231 (2000)
On the Design of Interface Agents for a DRT Transportation System Claudio Cubillos, Sandra Gaete, and Guillermo Cabrera Pontificia Universidad Cat´ olica de Valpara´ıso Escuela de Ingenier´ıa Inform´ atica Av. Brasil 2241, Valpara´ıso, Chile [email protected], [email protected], [email protected]
Abstract. This work tackles the design and development of interface agents present in a flexible transportation system devoted to the planning, scheduling and control of passenger trips. In particular, the agentoriented software engineering methodology (AOSE) PASSI was used. The interface agents devoted to interaction with customers and drivers are explained in detail and their prototypes are shown. The present work gives continuity to previous research in the field of transportation services and agent systems design.
1
Introduction
The need to cover more diffuse travel patterns, varying periods of low demand, city-peripheral journeys, as well as commuting trips often make conventional public transport systems unable to guarantee the level of service required to address the user needs. The use of Demand-Responsive Transport services (DRTS), where routes, departure times, vehicles and even operators, can be matched to the identified demand allows a more user-oriented and cost effective approach to service provision. On the other hand, software agents are defined as autonomous entities capable of flexible behavior denoted by reactiveness, pro-activeness and social ability [1]. Multiagent systems (MAS) consist of diverse agents that communicate and coordinate generating synergy to pursue a common goal. This higher level of abstraction has allowed agents to tackle the increasing complexity of nowadays open software systems. Under such an scenario, the agent paradigm has leveraged as an important modeling abstraction, in areas such as web and grid service, peer to peer and ambient intelligence architectures just to mention some cases. The present work describes the design of a multiagent system using a particular AOSE methodology called PASSI [2] for the passenger transportation under a flexible approach. It gives continuity to our past research [3] [4] on heuristics for solving scheduling of passenger trips. In particular the paper focuses in describing the design of the interface agents for the main actors; Customer and Driver, providing a practical case of GUI agents’ modeling with PASSI. A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 232–242, 2009. c Springer-Verlag Berlin Heidelberg 2009
On the Design of Interface Agents for a DRT Transportation System
2
233
Related Work
In general terms, Agent research on in the Intelligent Transportation Systems (ITS) domain has deserved an increasing interest. One of the ITS fields is in the Advanced Transportation Information System (ATIS), where Kase and Hattori [6] proposed the InfoMirror application that provides agent-based information assistance to drivers through car navigation systems or on-board PCs. A distributing route guidance system was presented by Adorni [7], which allowed dynamic route searching using the coordination capabilities of MAS. Another ITS field is Bus-holding control, which tackles the coordination of multiple lines of fixed-route buses and the different stops, seeking the global optimality. In 2001, Jiamin et al. [8] proposed a distributed bus-holding control approach in which a MAS negotiation between a Bus Agent and a Stop Agent was conducted based on marginal cost calculations. Deserves to be mentioned that in the above solutions no agent development methodology is used for specifying the system, mainly due to a lack of maturity of those methodologies at that time. Therefore, the aim of this work is to provide an experience in the design of multiagent systems using the PASSI methodology and to present the design of interface agents specific for the demand-responsive passenger transportation domain.
3
Flexible Public Transport Services
Demand Responsive Transport (DRT) services can be seen as an element of a larger intermodal service chain, providing local mobility and complementary to other conventional forms of transportation (e.g. regular buses and trams, regional trains). In this context, DRT provides a range of Intermediate Transport solutions, filling the gap between traditional public bus services and individual taxis. The DRT service can be offered through a range of vehicles including regular service bus, mini-bus, maxi-vans, buses and vans adapted for special needs and regular cars. The use of each vehicle type depends on the transport service to offer, the covered area and the target users. The aim aim is to meet the needs of different users for additional transport supply. The use of flexible transport services, where routes, departure times, vehicles and even operators, can be matched to the identified demand allows a more user-oriented and cost effective approach to service provision. The adaptation of the transport services to match actual demand enables cost savings to the operators, society and passengers. With respect to process implementation and management, the flexibility of the system is expressed along two main directions: on one hand, users of DRT systems must be provided with user-friendly instruments for accessing the services (such as information, reservation, query update) in several different flexible ways (the so called ”anywhere and anytime” access). On the other hand, the organization providing flexible services must be itself flexible, with the capability of managing dynamic relationships in a pool of transport resources (vehicles),
234
C. Cubillos, S. Gaete, and G. Cabrera
which may sometimes have to change to better adapt the transport supply to the dynamic demand.
4
The Agent-Based Transportation System
The system prototype was designed following the Process for Agent Societies Specification and Implementation (PASSI) which is made up of five models containing twelve steps in the process of building multi-agent. Please refer to [2] for a more detailed description on the whole PASSI methodology. Models were developed with the PTK (Passi Toolkit [14]) add-on for Rational Rose and was implemented over the Jade Agent Platform[9], which provides a full environment for agents to work. For a more detailed description on the agent architecture and the planning & scheduling mechanism please refer to [5] and [3]. In the following the general architecture will be explained for then detailing the interface agents in the next sections. As outlined in the PASSI section, the methodology starts capturing the system’s requirements through use cases, for then grouping them together to conform the agents. The diagram in Figure 1 shows part of the use cases and agents involved in the system. Due to space restrictions some of the supporting agents are expressed as actors. By starting from the transport operator side, we find the Vehicle, which is an interface agent (with a GUI) in charge of providing the monitoring of the routeschedule planned for the vehicle. In addition, it can inform the Driver about any changes to the initial schedule and can be used by him to inform any eventuality (e.g client no show, delay, detour, etc) that may happen regarding the trip and the customers. In particular its interface has been designed to work on-board the vehicle through a touch screen. This agent will be further detailed on a next section. The Schedule agent is the one in charge of managing the trip plan (workschedule) of the vehicle. In addition, the agent is also responsible of making trip proposals upon Planner request and in case of winning will have to update its actual plan to include the new trip. Upon changes (due to vehicle or client events) informed either by the Vehicle or the Planner agent, the Schedule agent will update the plan and reschedule the remaining requests. The Client is the second interface agent with a GUI, providing the connection between the end-user (Customer) and the transportation system. Through it, the Customer can request a trip by giving a description of the desired transportation service by means of a Trip Request Profile. In addition, using a Client Profile it is possible to create and manage personalized services profiles with diverse characteristics and preferences common to the different trips requested by the user. After a service has been contracted, the Customer can also communicate events to the system (e.g. a delay, a change on the agreed service, or simply cancel). In a similar way, the system can communicate with the Customer, informing him about any eventuality that may happen (e.g. a traffic jam or vehicle break down) which may imply a delay or change in the service to be provided. This agent will be also detailed on a next section.
On the Design of Interface Agents for a DRT Transportation System
235
Fig. 1. Portion of the Agents’ Identification Diagram
Other relevant agent is the Trip-Request, which acts as a proxy representing the Customer in the process of contracting a transportation service. In fact, the trip-request agent is involved in all the interaction of the Customer (through the interface agent) with the transportation system. Its activities regard the management of the client transportation requests, including any negotiation or selection of proposals coming from the Planner, together with processing any events generated by the Customer or by the system. As residing on a device with more processing power (such as a PC), this agent may have diverse degrees of autonomy for taking decisions on the trip proposal to choose and how to react when faced to eventualities. Finally, the Planner agent processes all the customers’ requests coming through their Trip-request agents. It initiates a contract-net (CNP) [15] with
236
C. Cubillos, S. Gaete, and G. Cabrera
the Schedule agents and manages all the arrived proposals. It is also in charge of managing events that may affect the trip services already contracted and scheduled. The remaining actors correspond to supporting-service agents or systems that interact with the diverse agents already detailed, such as the broker, responsible for the initial service matching, and the map, providing times and distances, among others. 4.1
The Client Agent
As stated before, the Client is an interface agent devoted to the Customer-System interaction. In principle, this trip-client assistant may reside on diverse devices (e.g PC, PDA, mobile phone) in order to allow a more flexible and pervasive access to the transportation system. In our prototype, has been developed a Client agent for PC, remaining the versions for more restricted devices as future work. In this sense, it is important to highlight that all the complex processing or decision-making (if delegated by the Customer) has been attached to the Trip-request agent in order to lightweight the Client (the interface agent). In the following Figure 2 a screenshot of the Client agent GUI is shown, detailing the tab that appears when initiating the request of a trip. In the ”Request Data” area, on the left, is asked all the information necessary to detail a transport service request under the demand-responsive considered scenario. This regards the date, the pickup and delivery points (addresses), the corresponding times and other specific information such as the required seats and diverse vehicle characteristics. On the right hand, the available transport services are deployed, showing for each selected service the covered area in terms of street intersections. The
Fig. 2. Client agent GUI showing the ”Request Data” tab in the ”Request Trip” menu
On the Design of Interface Agents for a DRT Transportation System
237
Fig. 3. Client agent GUI showing the ”Received Proposals Data” tab in the ”Request Trip” menu
services’ list can be imported from the system (on line) or from a local file. At the bottom, the Customer can send the trip request and save the services’ list. The following Figure 3 corresponds to a screenshot of the Client agent GUI detailing the ”Received Proposals Data” tab which appears as an answer after sending the request for a trip. In this form are displayed all the transportation alternatives found to be capable of performing the service. The list of alternatives is on the right-hand box and by selecting on each of them the left area (Selected Proposal Data) displays the details of such proposal. The data involved concerns the address and requested time for pickup and delivery. In addition, a time window is specified for the pickup and for the delivery in order to make more flexible the service and tackle possible differences with the original schedule. Other relevant data provided regards the vehicle ID and type, the required seats, together with diverse specific properties, such as the capacity, bicycle rack, shared/individual use, among others. It is important to mention that all the concepts involved in the specification of the services make part of a Domain Ontology specific for this transportation domain (for further details on the ontology please refer to [5]). The PASSI methodology used for the modeling considers a Task Specification step. In this activity the scope is to focus on each agents behavior, decomposing it into tasks which usually capture some functionality that conforms a logical unit of work. Therefore for each agent an activity diagram is developed, containing what that agent is capable of along the diverse roles it performs. In general terms, an agent will be requiring one task for handling each incoming and outgoing message. In the following Figure 4 a portion of the Task Specification Model for the Client Agent is depicted. The diagram shows six tasks that constitute the main Client
238
C. Cubillos, S. Gaete, and G. Cabrera
agent capabilities devoted to the process of requesting a transportation service. The SendQueryAvailableService task handles the request from the Customer to search for available services and triggers the ManageClientQuery task of the TripRequest agent which is in charge of requesting the Broker for possible transportations services available. These are returned by the SendActualAvailableService task of the Trip-request and is received by the ReceivingAvailableService task of the Client which processes and decodes the ACL message and forwards the services’ list to the ShowAvailableService task responsible for displaying the list in the proper form as already shown in Figure 3 right-hand box. The Customer, when making a trip Request Profile (see Figure 2), can browse on the available services (after loading them) in the right-hand area calling to the ShowAvailableService task or can send the request (by pressing the button) after filling the left-hand information, calling the SendTripRequest task. This Client’s task is responsible for sending the Request Profile to the Trip-Request, being handled by its ManageClientQuery task, which on its turn will forward the request to the Planner. The Trip-request agent will receive from the Planner the trip proposals coming from the different Schedule agents of the vehicles and its SendTripProposals task will send them to the Client. On its turn, the Client will receive and handle the proposals through its ShowTripProposals task, also responsible for displaying them on the appropriate form area as already shown on Figure 3. In this way, the Customer will be able to select the best alternative according to his preferences and will click the ”Confirm Selected Proposal” button on the GUI (see Figure 3 first button lower-right corner). This action will call the SendChosenProposal task of the Client responsible of forwarding the selected proposal to the Trip-request agent, which on its turn will forward it to the Planner, who is in charge of communicating the proposals’ acceptance/rejection to the diverse Schedule agents that made bids. 4.2
The Vehicle Agent
As mentioned earlier, the Vehicle agent constitutes an interface agent for the Driver - Transportation System interaction. From the Agents’identification Diagram of Figure 1, it is possible to see that the Vehicle is responsible for allowing the communication of incoming events to the Customer and of vehicle events to the transportation system. The following Figure 5 shows a screenshot of the Vehicle agent GUI, detailing the actual vehicle itinerary with expected times. On a first view, we can realize that the layout is minimalistic with simple shapes as buttons. This is because the Vehicle agent is intended to be on-board the transportation vehicle (car, van, maxi-taxi, etc.). Hence, the interfaces developed to be used in touch screens. On the left side appears the timetable, providing the expected times (e.g. 10:00, 10:30, 11:00 and so on) of the places to visit (either pickup or delivery points). These can be scrolled up or down with the square buttons in the lower part. On the right hand the interface is divided in three. A header on the top, showing the present date and time plus a square led that blinks when a change
On the Design of Interface Agents for a DRT Transportation System
239
Fig. 4. Part of the Task Specification Model for the Client Agent, showing the flow of tasks involved in the trip request processing
Fig. 5. Vehicle agent GUI showing the main screen with the itinerary
to the itinerary is carried out by the system and needs to be communicated to the driver. On the middle-right are shown the details of the entry selected on the left hand (the 10:00 in this particular case). It provides the destination address to reach, the time limit for departure in that place (10:05) for not arriving late to the next destination (at 10:30 in this case) and the number of passengers that go up and down in this stop. Additionally, it is provided the best route in order to arrive from the actual position of the vehicle to the required destination.
240
C. Cubillos, S. Gaete, and G. Cabrera
Finally on the footer part, the interface deploys three touch buttons; the first allows to confirm passengers presence at the stop, the second to inform a delay or anticipation with respect to the schedule and the third one to turn back to the main menu. The Task Specification Model for the Vehicle agent is depicted in Figure 6. This activity diagram contains five tasks that specify the labour carried out by the agent. The task ObtainEventData manages the notification of events received from the Driver when touching the Inform Event button of the Vehicle GUI (on Figure 5). The task handles the GUI processing of the notification at low-level, the notification details and forwards them to the SendEventNotification task. This task is responsible for translating the event notification and related data into an ACL message in alignment with the transportation domain ontology and finishes by sending to the Schedule agent the ACL message with the given event notification. The Inform Event button display another screen that allows to notify a delay due to a detour, a traffic jam, or a vehicle breakdown. On each stop the driver must confirm to the transportation system the presence or absence of the clients (Confirm Passengers button on Figure 5). For this, the driver’s passenger confirm action is managed by the ShowPassengersList task which is responsible for displaying other tab with a detailed list of the inbound and outbound passengers. After the list is displayed, in the case of inbound customers the driver can confirm the presence or absence on each particular case. This action is managed by the SendPassengerConfirmation task which takes the responsibility of taking the client details and sending the Schedule agent an ACL Message with the notification. Finally, the schedule through its SendTripUpdate task informs the Vehicle agent about changes in the itinerary. This messages are handled by the ShowDetailsOfUpdatedTrip task of the Vehicle agent. It is in charge of notifying the driver about a change, this by blinking the upper-right square led on
Fig. 6. Task Specification model for the Vehicle Agent
On the Design of Interface Agents for a DRT Transportation System
241
the screen (on Figure 5) together with updating the timetable in order to reflect the changes.
5
Conclusions
The presented work tackled the design description of interface agents using PASSI withing the context of a complete system for passenger transportation under a demand-responsive approach. The methodology used allowed an appropriate level of specification along its diverse phases. The focus was centered on the specification of the interface agents for the main actors involved with the system: Customers and Drivers; The Client and Vehicle Interface Agent prototypes were detailed providing an in-depth example of agent design & implementation using the AOSE PASSI. Additionally, the design enables showing how are articulated the Requirements model, providing a functional description of the system in terms of use-cases, the identification of agents’ roles through sequence diagrams and the identification of agents’tasks. Future work considers to perform usability tests on the interface agents in order to improve the quality of them, together with considering the development of client agents for mobile devices (e.g. palm or cellular phone). Other issue regards the system openness with the implementation of a high-level communication mechanism in order to provide a dynamic participation in such a system.
Acknowledgement This work is part of Project No. 209.746/2007 entitled ”Coordinaci´ on en una sociedad multiagente dedicada a la programaci´ on y control bajo ambiente din´ amico”, funded by the Pontifical Catholic University of Valpara´iso (www.pucv.cl).
References 1. Weiss, G.: Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence. MIT Press, Massachusetts, USA (1999) 2. Burrafato, P., Cossentino, M.: Designing a multiagent solution for a bookstore with the passi methodology. In: Fourth International Bi-Conference Workshop on AgentOriented Information Systems (AOIS 2002) (2002) 3. Cubillos, C., Crawford, B., Rodr´ıguez, N.: Distributed planning for the on-line diala-ride problem. In: Preparata, F.P., Fang, Q. (eds.) FAW 2007. LNCS, vol. 4613, pp. 124–135. Springer, Heidelberg (2007) 4. Cubillos, C., Rodriguez, N., Crawford, B.: A study on genetic algorithms for the ´ DARP problem. In: Mira, J., Alvarez, J.R. (eds.) IWINAC 2007. LNCS, vol. 4527, pp. 498–507. Springer, Heidelberg (2007) 5. Cubillos, C., Gaete, S.: Design of an agent-based system for passenger transporta´ tion using PASSI. In: Mira, J., Alvarez, J.R. (eds.) IWINAC 2007. LNCS, vol. 4528, pp. 531–540. Springer, Heidelberg (2007)
242
C. Cubillos, S. Gaete, and G. Cabrera
6. Kase, N., Hattori, M.: InfoMirror - Agent-based Information Assistance to Drivers. In: IEEE\IEEJ\JSAI Intelligent Transportation Systems Conference Proceedings, pp. 734–739 (1999) 7. Adorni, G.: Route Guidance as a Just-In-Time Multiagent Task. Journal of Applied Artificial Intelligence 10(2), 95–120 (1996) 8. Zhao, J., Dessouky, M., Bukkapatnam, S.: Distributed Holding Control of Bus Transit Operations. In: IEEE Intelligent Transportation Systems Conference Proceedings, Oakland - USA, pp. 976–981 (August 2001) 9. Bellifemine, F., et al.: JADE - A FIPA Compliant Agent Framework. C SELT Internal Technical Report (1999) 10. B¨ urckert, H., Fischer, K., et al.: TeleTruck: A Holonic Fleet Management System. In: 14th European Meeting on Cybernetics and Systems Research, pp. 695–700 (1998) 11. Fischer, K., M¨ uller, J.P., Pischel, M.: Cooperative Transportation Scheduling: An application Domain for DAI. Journal of Applied Artificial Intelligence 10 (1996) 12. Kohout, R., Erol, K., Robert, C.: In-Time Agent-Based Vehicle Routing with a Stochastic Improvement Heuristic. In: Proc. Of the AAAI/IAAI Int. Conf., Orlando, Florida, pp. 864–869 (1999) 13. Perugini, D., Lambert, D., et al.: A distributed agent approach to global transportation scheduling. In: IEEE/ WIC Int. Conf. on Intelligent Agent Technology, pp. 18–24 (2003) 14. PASSI Toolkit (PTK), http://sourceforge.net/projects/ptk 15. Smith, R.G., Davis, R.: Distributed Problem Solving: The Contract Net Approach. In: Proceedings of the 2nd National Conference of the Canadian Society for Computational Studies of Intelligence (1978)
Supporting Requirements Analysis in Tropos: A Planning-Based Approach Volha Bryl, Paolo Giorgini, and John Mylopoulos University of Trento, DIT, via Sommarive 14, Povo (TN) 38050, Italy [email protected], [email protected], [email protected]
Abstract. Software systems are becoming more and more part of human life influencing organizational and social activities. This introduces the need of considering the design of a software system as an integral part of the organizational and social structure development. Alternative requirements and design models have to be evaluated and selected from a social perspective finding a right trade-off between the technical and social dimension. In this paper, we present a Troposbased approach for requirements analysis, which adopts planning techniques for exploring the space of requirements alternatives and a number of social criteria for their evaluation. We describe the tool-supported analysis process with the help of a case study (the e-voting system), which is a part of a project funded by the Autonomous Province of Trento.
1 Introduction Unlike their traditional computer-based cousins, socio-technical systems include in their architecture and operation organizational and human actors along with software ones. Moreover, human, organizational and software actors in such systems rely heavily on rich inter-dependencies to other actors in order to fulfill their respective objectives. Not surprisingly, an important element in the design of socio-technical systems is the identification of a set of dependencies among actors which, if respected by all parties, will fulfill all objectives. Let’s make the problem more concrete. KAOS [5] is a state-of-the-art requirements elicitation technique that starts with stakeholder goals and through a systematic, toolsupported process derives functional requirements for the system-to-be and a set of assignments of leaf-level goals (constraints, in KAOS terminology) to external actors so that if the system-to-be can deliver the functionality it has been assigned and external actors deliver on their respective obligations, stakeholder goals are fulfilled. However, there are (combinatorially) many alternative assignments to external actors and the system-to-be. How does the designer choose among these? How can we select an optimal, or “good enough” assignment? What is an optimal assignment? The KAOS framework remains silent on such questions. Alternatively, consider the i∗ and Tropos frameworks for modeling and analysis of early requirements of agent-oriented systems [1,11]. In i∗/Tropos, goals are explicitly associated with external stakeholders and they can be delegated to other actors or the system-to-be. Or, they can be decomposed into subgoals that are delegated to other actors. In this setting, selecting a set of assignments A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 243–254, 2009. c Springer-Verlag Berlin Heidelberg 2009
244
V. Bryl, P. Giorgini, and J. Mylopoulos
(a) 1st alternative
(b) 2nd alternative
Fig. 1. Sample problem: two alternative models
is more complex because delegations can be transitive and iterative. “Transitive” means that actor A may delegate goal G to actor B who in turn delegates it to actor C. “Iterative” means that an actor A who has been delegated goal G, may choose to decompose it (in terms of an AND/OR decomposition) and delegate its subgoals to other actors. To illustrate the problem, consider the design task in Figure 1 where actor A1 has to achieve goal G, which can be refined into two subgoals G1 and G2 . The actor can decide to achieve the goal by itself or delegate it to actor A2 . In both cases, there are a number of alternative ways that can be adopted. For instance, A1 can decide to delegate to A2 the whole G (Figure 1a), or a part of it (Figure 1b). The diagrams follow Tropos modelling notation with circles representing actors, big dashed circles representing actors’ perspective, and ovals representing goals (interconnected by AND/OR-decomposition links). Social dependencies among actors for goals are represented by D-labelled directed links. Even for such a simple example, the total number of alternative requirements models is big, and a systemized approach and tool support for constructing and evaluating such networks of delegations would be beneficial. We are interested in supporting the design of socio-technical systems, specifically the design of a set of inter-actor dependencies intended to fulfill a set of initial goals. The support comes in the form of a tool that is founded on an off-the-shelf AI planner to generate and evaluate alternative assignments of actor dependencies to identify an optimal design. Specifically, our tool solves the following problem: given a set of actors, goals, capabilities, and social dependencies, the tool generates alternative actor dependency networks on the basis of the following steps, which may be interleaved or used inside one another: – Check problem-at-hand: (a) Analyze actor capabilities: check whether existing actors possess enough capabilities to collectively satisfy their goals; (b) Analyze actor inter-dependencies: check whether existing dependencies between actors allows them to fulfill all given goals. – Explore alternative dependency networks: (a) With the help of planning algorithms construct assignments of goals to actors that leads to the satisfaction of the actors’ goals; (b) Evaluate alternatives by assessing and comparing them with respect to a number of criteria, provided by the designer. The idea of casting the problem of designing a set of delegations for a socio-technical system into a planning one has already been presented in [4,3]. In addition, [3] proposes
Supporting Requirements Analysis in Tropos: A Planning-Based Approach
245
a preliminary approach to the evaluation of alternatives. However, what was missing in these works, is the systematic process of requirements analysis to support the design of socio-technical systems. In this paper, we present such a tool-supported process, which combines formalization and analysis of an organizational setting, the use of planning techniques [3], and a concrete set of optimality criteria to evaluate the produced requirements models. The approach is illustrated with the help of a case study involving an e-voting system, developed for the Autonomous Province of Trento. We also report on preliminary experimental results aimed to evaluate the scalability of the prototype tool. The rest of the paper is structured as follows. Sections 2 and 3 present the baseline for this research [3,4], namely, in Section 2 the details on applying planning techniques to requirements models construction are given, and in Section 3 criteria and procedures for evaluating those models are discussed. Section 4 presents the general process schema of the proposed requirements analysis approach. This is followed by the explanations on how the inputs to the process are organized and analyzed in Section 5. Section 6 introduces the e-voting case study and illustrates the whole approach on its basis, which is followed by summarizing remarks and discussion on the future work directions in Section 7.
2 Using Planning to Construct Requirements Models The task of constructing delegation networks can be framed as a planning problem: selecting a suitable design corresponds to selecting a plan that satisfies actors’ and organizational goals [3]. In general, AI (Artificial Intelligence) planning [10] is about automatically determining the course of actions (i.e. a plan) needed to achieve a certain goal where an action is a transition rule from one state of the world to another. A specification language is required to represent the planning domain, i.e. the initial and the desired states and the actions. The following predicates are used to formally represent the initial and the desired states of the world (i.e. the organizational setting). The predicates take variables of three types: actors, goals and goal types. To typify goals, type(goal, g type) predicate is used. Actor capabilities are described with can satisfy(actor, goal) and can satisfy gt(actor,g type) predicates meaning that an actor has enough capabilities to satisfy either a specific goal, or a goal of a specific type. Dependencies among actors are reflected by can depend on(actor, actor) and can depend on gt(actor, actor, g type) predicates, which mean that one actor can delegate to another actor the fulfilment of any goal or, in the latter case, any goal of a specific type. Predefined ways of goal refinement are represented by and/or decomposition(goal, goal, ..) predicates. Initial actor desires are represented with wants(actor, goal) predicate. When a goal is fulfilled satisfied(goal) predicate becomes true for it. In Figure 2a the example presented in Figure 1 (hereafter referred to as sample problem) is formalized using the above described predicates.
246
V. Bryl, P. Giorgini, and J. Mylopoulos A1, A2 – t actor G, G1, G2 – t goal (can satisfy A1 G) (can satisfy A1 G1) (can satisfy A1 G2) (can satisfy A2 G) (can satisfy A2 G1) (can satisfy A2 G2) (can depend on A1 A2) (and decomposition2 G G1 G2) (wants A1 G)
DELEGATES A1 A2 G SATISFIES A2 G (b) Plan for 1st alternative AND DECOMPOSES A1 G G1 G2 SATISFIES A1 G1 DELEGATES A1 A2 G2 SATISFIES A2 G2 (c) Plan for 2nd alternative
(a) Formalization Fig. 2. Sample problem: formalization and plans
A plan, which is constructed to fulfill actors and organizational goals, comprises the following actions: (a) Goal satisfaction, (b) Goal delegation, and (c) Goal decomposition. Actions are described in terms of preconditions and effects, both being conjunctions of the above mentioned predicates and/or their negations. If a precondition of an action is true in the current state, then the action is performed; as a consequence of an action, a new state is reached where the effect of the action is true. When the problem domain and the problem itself are formally represented, a planner is used to produce a solution. Having analyzed a number of available o-the-shelf planners (see [4] for the details), we have chosen LPG-td [7], a fully automated system for solving planning problems, which supports PDDL 2.2 specification [6], for implementing our planning domain. In Figure 2b-c two plans for the sample problem generated by LPG-td are shown. A set of experiments, which we do not report here due to the space limits, were conducted with LPG-td to assess the scalability of the approach. The key point of the experiments was to identify the complexity limits (in terms of number of goals, actors and actor dependencies, depth of goal decomposition trees, etc.) which the planner could handle. The obtained results justify the use of planning in the requirements engineering domain as, according to our experience, requirements models of real life case studies stay within the complexity limits which our planning approach is able to manage. In [3] we have presented P-Tool, an implemented prototype to support the designer in the process of exploring and evaluating alternatives. The tool has the interface for the input of actors, goals and their properties. LPG-td is built in the tool, and is used to generate alternative requirements structures, which are then represented graphically using Tropos notation.
3 Evaluating Requirements Models After a requirements model has been constructed, it can be evaluated from the two perspectives, global (i.e. the perspective of the designer), and local (i.e. the perspective of an individual actor). The perspective of the designer will normally include a number of non-functional requirements, which he wants to be met. E.g., these might be security
Supporting Requirements Analysis in Tropos: A Planning-Based Approach
247
concerns, efficiency, maintainability, user friendliness, etc. Evaluation is either quantitative, or qualitative, depending on whether one can “measure” the requirement in terms of numbers, or there exists just a relative scale to compare two models. An example of quantitative criteria would be a “fair”, or homogeneous distribution of workload among the actors, with a variance as a measure of it. An example of qualitative criteria could be user friendliness, meaning that the expertise of the designer allows him to say that organizing the process in this way is “more user friendly” than in this other. An example of a local evaluation criterium we consider in this paper, is related to workload distribution. We assume that each actor, either human or software, wants to minimize the number and complexity of goals it/he/she needs to satisfy. Complexity of a goal for an actor measures the effort required to achieve the goal. Complexity has to be defined explicitly for leaf goals, i.e. for those goals that could be assigned to actors that have capabilities to satisfy them. There is no need to define or calculate complexity for a goal that is to be further decomposed and delegated. Complexity is “local” in a sense that the same goal can have different complexity values for different actors. For each actor there is a maximum complexity it can handle, i.e. the sum of complexity values for all the goals this actor is assigned should be less than some predefined threshold, namely, maximum complexity. If this condition is violated the actor might be willing to deviate from the imposed assignment. After the complexity and maximum complexity values are defined, the evaluation procedure, which preliminary version was presented in [3], is as follows. 1. A plan P is generated by the planner. 2. Plan complexity values for each actor are calculated, by summing up the complexity values for all the goals this actor is assigned. 3. Actors whose plan complexity values are greater than the corresponding maximum complexity values are identified. 4. One of these actors is selected, namely, actor amax which has the maximum difference δ between plan complexity and maximum complexity values. 5. A subset of actions Pdev ⊂ Pamax is formed with the total cost greater or equal to δ , where Pamax denotes those actions of P in which amax is involved. 6. The definition of the planning problem is changed in order to avoid the presence of actions contained in Pdev during the next planning iteration. 7. The procedure re-starts with the generation of a next plan.
4 Requirements Analysis: A General Schema Our proposal is to structure the requirements analysis process to support a designer in constructing and evaluating requirements models. The general schema of the process is presented in Figure 3. A preliminary description of an organizational setting, which is provided by a designer in terms of actors, goals and social relations, is analyzed and iteratively improved so as to output a model that guarantees the fulfillment of stakeholder goals and is good-enough with respect to a number of criteria. In the following we give details on the process steps and their interrelations. Most of the process steps can be
248
V. Bryl, P. Giorgini, and J. Mylopoulos
Fig. 3. Requirements analysis process: a general schema
automated, however, the presence of a human designer is inevitable the design process for socio-technical systems can be supported by tools but cannot be automated. As a first step, it is checked whether there exist at least one assignment of goals to actors that leads to the satisfaction of top-level goals. Input checker analyzes the organizational setting, detects inconsistencies, and proposes possible improvements, which then are either approved, or rejected, or modified by the designer. In particular, it is checked whether available actors possess enough capabilities to collectively satisfy their goals, and whether the relationships between actors permit this to happen. To analyze actor capabilities means to check that for each goal it is possible to find an actor capable of achieving each of its AND-subgoals or at least one of its OR-subgoals. To analyze actor interdependencies means to check whether a goal can be delegated from an actor who wants to achieve it to an actor who is capable of achieving it. Given a network of delegations between the actors, it is checked whether there exists a path between two actors. In Section 5, we give details on analyzing and dealing with missing capabilities, while the second kind of analysis is not covered in the paper. After the input is checked, the first possible alternative is generated by the Planner component, which exploits AI planning algorithms to search for a solution as described in Section 2. An alternative generated by the planner is then assessed by the Evaluator with respect to a number of criteria. These criteria are defined by the designer and refer to the optimality of the solution either from a global perspective (e.g. assessing the overall security or efficiency), or from the local perspectives of stakeholders (e.g. assessing the workload distribution). Evaluation criteria and procedures were discussed in Section 3. If evaluation reveals that an alternative is not acceptable, then the Evaluator provides feedback to the Planner in order to formulate constraints for the generation of the next alternative. If no further alternative can be generated, the current description of an organizational setting is changed according to the constraints identified by the Evaluator, and then is analyzed by the Input checker, and so on, iteratively.
Supporting Requirements Analysis in Tropos: A Planning-Based Approach
249
Note that the output of the evaluation process needs to be approved by a human designer. User evaluation interface presents the selected alternative to the designer together with the summarized evaluation results. Another task of this component is to provide the designer with the interface for giving his feedback on why the selected alternative does not satisfy him. On the basis of this feedback the constraints for the generation of the next alternative are formulated and forwarded to the Planner. The result of the application of the approach is a new requirements model, which is, ideally, optimal or, in practice, good-enough with respect to all the local and global criteria, and is approved by the designer. After obtaining one satisfiable alternative it is possible to repeat the process to generate others, reusing already obtained constraints.
5 Input Analysis The input to the proposed requirements analysis process is provided by the designer, and can be logically divided into the following two categories. Firstly, it is an organizational setting description, i.e. actors, their goals and capabilities, dependencies among actors, possible ways of goal refinements (decompositions of a goal into AND/OR-subgoals). The second category contains evaluation criteria, either qualitative or quantitative, e.g. preferences or cost bounds, as discussed in Section 3. As it was stated in Section 4, the description of an organizational setting can be analyzed with respect to capabilities of available actor and actor connectivity. In this section the process of analyzing actor capabilities is explained, while actor connectivity analysis is outside the scope of this paper. Firstly, a “capability tree” for each high-level organizational goal is constructed. A “capability tree” is based on a goal model [1], which organizes goals in a tree structure reflecting AND- and OR-decompositions. In addition, a list of actors capable of achieving a goal is associated with each leaf node of a tree. Nodes with no associated actors are marked as unsatisfiable with the goal which causes the problem included in the label (unsat : goal). Otherwise, a node is marked as satisfiable (sat). After that, labels are propagated bottom-up to the root goal according to the following rules. If all OR-subgoals or at least one AND-subgoal of a goal is marked unsat, then the goal is also marked unsat, and sat otherwise. When an unsat label is propagated, the goals that cause unsatisfiability problem are accumulated together; at the end the label might look as unsat : (G1 or G2 ) and G3 , meaning that to satisfy the root goal we need to find the ways to satisfy G3 and either G1 or G2 . After a capability tree has been built and sat/unsat labels propagated, it becomes clear whether the root goal can be satisfied, and if not, what are the missing capabilities. There are two ways to deal with missing capabilities: – Add a new capability to an existing actor. Such a decision could be based on the actual capabilities of this actor. Namely, if this actor is already capable of achieving one or several goals of the same type, it is likely that it could manage the new goal as well. – If there is no way to add the missing capability to one of the existing actors, a new actor might be introduced.
250
V. Bryl, P. Giorgini, and J. Mylopoulos
6 E-Voting Case Study The focus of the case study is on modelling requirements for an electronic based voting system that is to be introduced in the Autonomous Province of Trento by the next provincial elections (to be held in 2008). The Province is funding the ProVotE project [9], which has the goal of providing a smooth transition from the paper based voting system to new technologies. The project includes partners from the public administration, research centers and academia, and local industries. 6.1 Description Voting in Italy consists of the following stages [8]: (a) Identification and registration of a voter at a polling station; (b) Casting a vote; (c) Counting votes and tabulating the results; (d) Transmission of the results to the offices responsible for data aggregation; (e) Sum and proclamation of the elected representatives. Introducing the new technologies in all these stages not only changes the way in which votes are collected and processed, but also roles and responsibilities of the actors involved. One of the ProVotE activities was the extensive UML modelling of existing paper based voting procedures [8] in order to provide a baseline for the definition of new procedures. Obviously, there are many ways in which the paper based system might be transformed into an e-voting system, and thus, many alternative configurations should be analyzed and compared. However, UML modelling approach, being process-oriented, does not provide support for such an analysis. To complement UML modelling in tackling the above problems, Tropos modelling approach has been used, which provides a clear visual representation of the organizational setting (actors, goals, actor dependencies) combining the perspectives of different stakeholders1.
Fig. 4. Case study: conduct elections in a section 1
See [2] for the discussion on why and how UML and Tropos modelling approaches are applied together in ProVotE.
Supporting Requirements Analysis in Tropos: A Planning-Based Approach
251
Fig. 5. Case study: transfer processed data
Example 1. In Figure 4 the diagram representing the viewpoint of the polling station staff on the voting process is shown2. As presented in the diagram, the root level goal, conduct election in a section, is decomposed into the three AND-subgoals, two of which are then further decomposed. Some of the subgoals (provide access to vote casting environment, and transfer processed data) can be satisfied by the polling station staff. For the achievement of other goals the polling station staff depends on the other actors, either organizations, or humans, or software systems, e.g. the goal register voters is delegated to the registration system actor. Example 2. The second example concerns the goal transfer processed data, and the president of the section, the leading member of the polling station staff, who is responsible for satisfying this goal, see Figure 5. Elections are conducted with the help of several voting machines installed in standard voter cabins in each polling station. A voter expresses his/her choice with the help of a touch-screen, and then a ballot is printed, approved by the voter and stored inside a machine. At the end of an election day, data from voting machines are collected on the special purpose USB keys, one per machine, and printed paper ballots are extracted from the machines (more details are given in [8]). Thus, there are three pieces of data that could be transferred from the polling station to the Electoral office, namely, USB keys, electronic data these keys contain, and paper ballots. 6.2 Illustrating the Process Capability analysis. Figure 4 depicts a capability tree for conduct elections in a section goal, in which two leaf goals are assigned sat labels, namely, provide access to 2
For the moment sat/unsat labels should be ignored, the necessary explanations will be given in Section 6.2.
252
V. Bryl, P. Giorgini, and J. Mylopoulos Table 1. Case study: plans and their evalaution # 1 2
3
4
5
Description Actor : Workload Depicted in Figures 5 and 6a President: 20 President delegates the goal of preparing President : 15 USB keys for sending to Secretary. Secretary : 5 Designer decides that to maintain the required security level all the data means should be transferred to the Electoral President : 35 office. Transfer all data means goal is Secretary : 5 adopted, Secretary is in charge of preparing USB keys for sending, President takes care of all the other subgoals. The goal of preparing paper ballots for President : 25 sending is delegated to Secretary. Secretary : 15 No plan can be generated, so actors capabilities are analyzed, and it is inferred that President : 15 Scrutineer can satisfy a goal of extracting Secretary : 15 paper ballots from voting machines. This Secretary : 10 goal is delegated to Scrutineer, see Figure 6b.
Who deviates President Designer
President
President
–
vote casting environment and transfer processed data, meaning that polling station staff possesses enough capabilities to satisfy both of them. After sat/unsat labels are propagated, it becomes clear that the goal conduct elections in a section cannot be satisfied due to the fact that the polling station staff is not able (or need extra support) to register voters, count ballots, and aggregate the results centrally3 . To address the missing capability problems, the following steps are performed. Firstly, the capability to aggregate the results centrally is added to the existing actor, Electoral office; in the initial input setting this data could be omitted just by mistake, and such kind of input analysis helps to correct the errors. Secondly, two new actors are introduced to deal with register voters and count ballots goals, registration system and voting machine, respectively. These new actors are the parts of the new e-based voting system, specifying the requirements to which is what we aim at. Planning and evaluation. In Table 1 the iterations of the planning and evaluation procedure applied to the case study are presented. When evaluating the workload, it is assumed that the goals send data to Electoral office via PC and prepare USB keys for sending are considered to be not very complex goals with the complexity value equal to 5 units for each of the two goals. The other three leaf goals are considered to be more complex with the complexity value equal to 10 units. For the sake of simplicity the complexity values for the same goal are the same for all actors. Maximum complexity values are the same for all the actors and are equal to 15 units. 3
Extensions of unsat labels (which accumulate the goals that cause unsatisfiability problem as it was described in Section 5) are omitted for the sake of presentation simplicity.
Supporting Requirements Analysis in Tropos: A Planning-Based Approach
253
OR DECOMPOSES PRES TR TRALL TREP TREK AND DECOMPOSES PRES TRALL TRE TRK TRP OR DECOMPOSES PRES AND DECOMPOSES PRES TR TRALL TREP TREK TRP EXTRACTP PREPAREP AND DECOMPOSES PRES DELEGATES PRES SCRU EXTRACTP TREK TRE TRK SATISFIES SCRU EXTRACTP AND DECOMPOSES PRES DELEGATES PRES SECR PREPAREP TRE UPLOADONK SENDE SATISFIES SECR PREPAREP SATISFIES PRES UPLOADONK AND DECOMPOSES PRES SATISFIES PRES SENDE TRE UPLOADONK SENDE AND DECOMPOSES PRES SATISFIES PRES UPLOADONK TRK UPLOADONK PREPAREK SATISFIES PRES SEND SATISFIES PRES PREPAREK AND DECOMPOSES PRES (a) Initial Plan TRK UPLOADONK PREPAREK DELEGATES PRES SECR PREPAREK SATISFIES SECR PREPAREK (b) Final Plan Fig. 6. Transfer data: the initial and the final plans
7 Conclusions and Future Work In this paper we have proposed a structured Tropos based requirements analysis process which focuses on exploring and evaluating alternative requirements models. Planning techniques are adopted for constructing alternative dependency networks, which are then evaluated with respect to a number of criteria. Evaluation criteria represent either a designers point of view, or assess the models from the local perspectives of organizational actors. The approach is illustrated with the help of the case study taken from the ongoing project, which is about introducing e-based voting system in Autonomous Province of Trento. The priority among the possible future work directions is on further implementation and testing the tool which supports all steps of the design process. Other important research directions include the further development of evaluation metrics, with a focus on local evaluation metrics; elaborating on the input analysis to address all the situations in which no solution is available; exploiting advances in AI planning by incorporating some of the evaluation metrics into the planning process; further experiments regarding the scalability and usability of the proposed approach.
Acknowledgements This work has been partially funded by EU Commission, through the SENSORIA and SERENITY projects, by MEnSA-PRIN project, and also by the Provincial Authority of Trentino, through the MOSTRO project.
254
V. Bryl, P. Giorgini, and J. Mylopoulos
References 1. Bresciani, P., Giorgini, P., Giunchiglia, F., Mylopoulos, J., Perini, A.: Tropos: An agentoriented software development methodology. JAAMAS 8(3), 203–236 (2004) 2. Bryl, V., Ferrario, R., Mattioli, A., Villafiorita, A.: Evaluating Procedural Alternatives in an e-Voting Domain: Lesson Learned. Technical Report DIT-07-005, University of Trento (2007) 3. Bryl, V., Giorgini, P., Mylopoulos, J.: Designing cooperative IS: Exploring and evaluating alternatives. In: Meersman, R., Tari, Z. (eds.) OTM 2006. LNCS, vol. 4275, pp. 533–550. Springer, Heidelberg (2006) 4. Bryl, V., Massacci, F., Mylopoulos, J., Zannone, N.: Designing security requirements models through planning. In: Dubois, E., Pohl, K. (eds.) CAiSE 2006. LNCS, vol. 4001, pp. 33–47. Springer, Heidelberg (2006) 5. Dardenne, A., van Lamsweerde, A., Fickas, S.: Goal-directed requirements acquisition. Science of Computer Programming 20, 3–50 (1993) 6. Edelkamp, S., Hoffmann, J.: Pddl2.2: The language for the classical part of the 4th international planning competition. Technical Report 195, University of Freiburg (2004) 7. LPG Homepage. LPG-td Planner, http://zeus.ing.unibs.it/lpg/ 8. Tiella, R., Villafiorita, A., Tomasi, S.: Specification of the Control Logic of an eVoting System in UML: the ProVotE experience. In: CSDUML 2006 (2006) 9. Villafiorita, A., Fasanelli, G.: Transitioning to eVoting: the ProVotE project and Trentino’s experience. In: EGOV 2006 (2006) 10. Weld, D.S.: Recent Advances in AI Planning. AI Magazine 20(2), 93–123 (1999) 11. Yu, E.S.-K.: Modelling strategic relationships for process reengineering. PhD thesis, University of Toronto (1996)
Towards Method Engineering for Multi-Agent Systems: A Validation of a Generic MAS Metamodel Graham Low1 , Ghassan Beydoun3, Brian Henderson-Sellers2, and Cesar Gonzalez-Perez1,2 1
2
School of Information Systems, Technology and Management, University of New South Wales, Sydney, Australia Faculty of Information Technology, University of Technology of Sydney, Sydney, Australia 3 Faculty of Informatics, University of Wollongong, Wollongong, Australia
Abstract. It is a widely accepted premise that no single methodology can be suitable for all Multi-Agent System (MAS) software projects. This premise is playing a significant role in the appearance of new methodologies at an increasing pace. To effectively harness the software engineering knowledge of methodologies, method engineering is an appealing approach. It focuses on project-specific methodology construction from existing method fragments and it requires a generic product-focussed metamodel to serve as a representational infrastructure to unify existing methodologies into a single specification. As part of its ongoing validation towards method engineering for MAS development, we present our generic metamodel and illustrate in this paper its wide applicability with respect to 10 methodologies extant agent-oriented. This validation provides further evidence to support the use of our metamodel towards the construction of situated MAS methodologies.
1 Introduction Multi Agent Systems (MAS) are a new class of distributed parallel software systems that have proved effective in the following core tasks: automating management of information within businesses (e.g. computer network management applications [1]), building computational models of human societies to study emergent behaviour [2–4] and building cooperative distributed problem solving [5,6]. The building blocks of a MAS are intelligent, autonomous and situated software entities: agents. The agent, the concept of agency and the MAS abstractions offer a promise of making software systems easier to embed within our daily lives as suggested in [7]. However, it is widely acknowledged that for agents to be accepted into the mainstream of software development, suitable development methodologies are required. As a testimony to this, agent-oriented methodologies are being published at an increasing rate. We have identified more than a dozen, examples being Gaia [8], Adelfe [9], Prometheus [10] and PASSI [11]. Formal arguments refuting that a single methodology is sufficient, regardless as to how well thought out it might be, have recently appeared [7]. We support such arguments, in our work, rather than the suggestions that combining existing methodologies A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 255–267, 2009. c Springer-Verlag Berlin Heidelberg 2009
256
G. Low et al.
into one comprehensive methodology is the answer e.g. [12]. We believe the latter will prove to be impossible, because the sets of assumptions underlying each methodology are likely to be inconsistent and irreconcilable. In [13], we proposed an alternative to empower MAS software developers to create methodologies from existing (method) fragments (self-contained components); this activity is known in software engineering as method engineering [14]. Method engineering approaches have been successful in object-oriented development due to widely accepted modelling languages and the constructs of OO software systems and development processes e.g. [15]. With the objective of applying method engineering for developing a MAS, in [13] we synthesised and introduced a process-independent metamodel for an agent-oriented modelling language to describe software components of any MAS. This metamodel is the first to focus on conceptual and ontological underpinnings rather than being constrained for use in a single methodological approach. It is widely applicable, without any constraining preequisites; it will be able to be used to describe work products required in most methodologies. At the system level, we did not make any assumptions about agents beyond their essential properties: autonomy, situatedness and interactivity. Any other non-definitional agent characteristic, visible at the system level, suggests a methodology-specific feature. For instance, adaptivity is non-definitional. Adelfe’s [16] adaptive system design requires learning agents. Thus, some concepts in Adelfe’s metamodel are too specific to be included in our metamodel. At the agent level, we intend our metamodel to be comprehensive enough to describe any type of agent (reactive, planning). In other words, to describe a given methodology, not all aspects of our metamodel will be needed. Our metamodelling approach is iterative and comprises two iterative phases: synthesis and validation. To validate our metamodel in [13], we verified that it can be successfully used to obtain metamodels underlying Islander [17] and TAO [18]. We sketched how our metamodel can generate both of them. That preliminary validation reinforced our case for method engineering and constituted early evidence that our method engineering proposal for MAS development is plausible and it is one step closer to being realised. In this paper, we undertake further validation of our metamodel by ensuring that the metamodel supports the key AOSE concepts considered by [19] when evaluating 10 extant and well-known methodologies: Gaia [8], Tropos [20], Adelfe’s [16], Prometheus [10], PASSI [11], MAS-CommonKADS [21], MASE [22], RAP [23] MESSAGE [24] and Ingenias [25]. The rest of this paper is organised as follows: In Section 2, we present our metamodel and how it was synthesized. In Section 3, we present our validation while in Section 4 we conclude, pinpointing where the metamodel can be improved as a result of the evaluation and describe future work.
2 MAS Metamodel for Method Engineering The full details of the synthesis of our metamodel is presented in [13]. In this section, we briefly overview that synthesis process and then present our actual metamodel before we undertake its validation in Section 3.
Towards Method Engineering for Multi-Agent Systems
257
Table 1. Design-time concepts and their definitions Term Action Specification Agent Definition Convention Environment Statement Facet Action Specification Facet Definition Functional Requirement Message Action Specification Message Schema Non-Functional Requirement Ontology Ontology Aggregation Ontology Concept Ontology Relationship Ontology Specialisation Performance Measure Plan Specification Requirement Role System Task
Definition Specification of an action, including any preconditions and postconditions. Specification of the initial state of an agent just after it is created. Rule that specifies an arrangement of events expected to occur in a given environment. A statement about the environment. Specification of a facet action in terms of the facet definition it will change and the new value it will write to the facet. Specification of the structure of a given facet, including its name, data type and access mode. Requirement that provides added value to the users of the system. Specification of a message action in terms of the message schema and parameters to use. Specification of the structure and semantics of a given kind of messages that can occur within the system. Requirement about any limits, constraints or impositions on the system to be built. Structural model of the application domain of a given system. Whole/part relationship between two ontology concepts. Concept included in a given ontology. Relationship between ontology concepts. Supertype/subtype relationship between two or more ontology concepts. Mechanism to measure how successful the system is at any point in time. An organised collection of action specifications. Feature that a system must implement. Specification of a behavioural pattern expected from some agents in a given system. Final product of a software development project. Specification of a piece of behaviour that the system can perform.
2.1 Overview of Our Metamodel Synthesis We ensured that our metamodel is a consensual picture of what a MAS looks like. The resultant metamodel covered as many features of a MAS as possible. In its synthesis, the behavioural characteristics of agents in a wide range of MASs were considered. The metamodel describes the behaviour of any agent. We anticipated that any methodology can be successfully generated using our metamodel. We aimed to maximise coverage (inclusion of as many MAS concepts as possible) and generality (wide acceptance and
258
G. Low et al.
familiarity to methodologists). To construct our generic metamodel, we decided on the set of concepts that will be used, describing entities in any MAS and the relationships amongst them. This set of concepts and their definitions was rooted in the existing literature related to MAS and MAS methodologies. Because of the sheer size of this literature, we could not be sure that the initial set of examined references was comprehensive. We continuously considered additional references during the metamodelling process. This necessitated that our metamodelling process be iterative. This process consisted of the following steps undertaken iteratively: Step 1: Decide on the set of general concepts relevant to any MAS and its model. See Tables 1–2 for a list of concepts from this step. Step 2: Decide on definitions worth considering. Step 3: Reconcile differences between definitions where possible, giving hybrid definitions. Step 4: Designate the chosen concepts into two sets: run-time concepts and design-time concepts. Each set has two scopes: system-related or agent internals related scope. Step 5: Identify the relations within both the design-time and run-time sets produced in Step 4, to produce the actual metamodel. These steps did not depend on any single software development methodology. Table 2. Run-time concepts and their definitions Term Action Agent Belief Desire Environment Environment History Environment Statement Event Facet Facet Action Facet Event Goal Intention Message Message Action Message Event Obligation Plan
Definition Fundamental unit of agent behaviour. A highly autonomous, situated, directed and rational entity. An environment statement held by an agent and deemed as true in a certain timeframe. An environment statement held by an agent, which represents a state deemed as good in a certain timeframe. The world in which an agent is situated. The sequence of events that have occurred between the environment start-up and the present instant. A statement about the environment. Occurrence of something that changes the environment history. Scalar property of the environment that is expected by the agents contained in it. Action that results in the change of a given facet. Event that happens when the value of a facet changes. Ultimate desire. A committed desire. Unit of communication between agents, which conforms to a specific message schema. Action that results in a message being sent. Event that happens when a message is sent. Behaviour expected from an agent at some future time. An organised collection of actions.
Towards Method Engineering for Multi-Agent Systems
259
3 Our Generic Metamodel Our metamodel is complete as far as describing the internal structure of single agents is concerned (according to our three definitional properties of agents). However, not all concepts in the metamodel have to be used by a given methodology. Some issues are left to the methodology or the developers, e.g. how plans are generated and dumped, how beliefs are maintained and shared, verification and validation of the system. In connecting all concepts into one coherent metamodel, we omitted all relations which are too specific, relating to some special kind of agent. We also ensured that the set of terms was self-contained, that is, concepts may only depend on each other in this set. The metamodel is summarized below (for full details see [13]). It has two layers: design-time and runtime layers and can have one of two scopes: system related or agent related. The metamodel is presented in four diagrams to clearly group classes into the four different areas of concern: design-time system related classes, runtime system-related (environment) classes, design time agent-internals classes and run-time agent-internals classes. Fig. 1 shows the classes (and associated relationships) of the metamodel that are directly related to the description of a MAS, i.e. design-time system-related classes. These classes are concerned with features that can only be perceived by looking at the whole system at design time: Role, Message schema, Task, Agent Definition, Ontology plus environment access points.
Fig. 1. System related design-time classes after [13]
260
G. Low et al.
Fig. 2. Run-time, environment-related classes after [13]
Fig. 3. Agent-internals design-time classes after [13]
Fig. 2 shows the classes related (and associated relationships) to the environment in which agents “live”, that is, run-time environment-related classes. These are concerned with MAS features that exist only at runtime in the environment: Environment History, Event together with system access points. Fig. 3 shows the classes (and associated relationships) related to the agent internals at design time: Plan Specification and Action Specification. Finally, Fig. 4 shows the classes related (and associated relationships) to agent internals at run-time: Plan, Action, Desire (and Belief) and Intention.
Towards Method Engineering for Multi-Agent Systems
261
Fig. 4. Agent-internals design-time classes after [13]
4 Metamodel Validation We are concerned with external consistency and completeness of the metamodel [as discussed earlier, internal consistency is ensured during its creation]. External validation is designed to ensure external consistency with existing methodologies and ensure that the metamodel can generate all prominent methodologies. Generally, internal consistency is a strong indicator of external consistency as demonstrated and highlighted in Beydoun et al. [26]. In our validation, we directly undertook an external consistency check against the explicit and implicit metamodels of methodologies available in the literature. We validated our metamodel using TAO and Islander as reported in [13]. In this section, we determine if our metamodel supports the key AOSE concepts considered by [19] when evaluating 10 extant and well-known methodologies (shown in Table 3). Tran and Low [19, 27] developed an evaluation framework by identifying and integrating the evaluation criteria from various feature analysis frameworks, including those for assessing conventional system development methodologies [28-30], Object Agency Inc. [31] and those for evaluating AOSE methodologies [32-35]. The former category provides a well-established list of generic system engineering features to be considered, while the latter presents various agent-oriented and MAS-specific aspects
Fig. 5. Comparison of Concepts
262 G. Low et al.
Towards Method Engineering for Multi-Agent Systems
263
Table 3. Support of Concepts by Generic Metamodel Concepts System goal System task/behaviour Use case scenario Role Domain conceptualization Agent goal/task Agent-role assignment Agent belief/knowledge Agent capability/service Agent plan/reasoning rule/problem solving method Agent percept/event Agent architecture Agent acquaintance Interaction protocol Content of exchanged messages Inter-agent contract/commitment System architecture Organizational structure/inter-agent social relationship Environment resource/facility Environment characterization Agent aggregation relationship Agent inheritance relationship Agent instantiation Agent instances deployment Ontology Mobile agents
Support by FAML metamodel Not supported Task class Not supported Role class Ontology classes Task class Role Specification between Role and Agent Definition classes Belief, Intention and Desire classes Plan class Event class Agent Definition class Role Relationship classes Not supported Message class Message class Not supported Organization: not supported Inter-agent social relationship: handled via roles Environmental level classes Environmental level classes Extension of Agent Definition class Extension of Agent Definition class Agent class Agent Definition class Ontology classes Not supported
for assessment. They also added several evaluation criteria that are not included in existing frameworks, e.g. “support for mobile agents” and “support for ontology”. They conducted a survey to substantiate the relevance of the proposed criteria and to make the necessary refinements to the criteria specification. The framework consists of: – Process Related Criteria: evaluating an AOSE methodology’s development process. – Technique Related Criteria: assessing the methodology’s techniques to perform development steps and/or to produce models and notational components. – Model Related Criteria: examining the capabilities and characteristics of the methodology’s models and notational components including the concepts represented by the models, the quality of notational components and the agent characteristics supported by the models. – Supportive Feature Criteria: evaluating a variety of high-level methodological capabilities.
264
G. Low et al.
This structure highlights the framework’s attention to all three major components of a system development methodology, i.e. process, models and techniques. Tran and Low [19] used the feature analysis technique to evaluate 10 well known AOSE methodologies. The concepts related criteria are particularly relevant to our validation of the generic metamodel. Two additional criteria included in supportive-feature criteria are considered: ontology and support for mobile agents. The “Concepts” criterion in Table 3 evaluates the modelling capability of an AOSE methodology by examining which concepts the methodology’s models are capable of capturing/representing both at design time and run time. A list of “standard” AOSE concepts was used as an assessment checklist. If a methodology supports the modelling of a particular concept, the name of the model or notational component (i.e. diagram) that represents that concept is indicated. If a concept appears in many different models or diagrams within the methodology, the model or diagram that represents the concept as the principal modelling element is specified. Table 4 examines the support of the generic.
5 Discussion, Summary and Future Work In [13] we provided a process-independent metamodel for an agent-oriented modelling language to describe software components of any MAS and showed how our metamodel can be refined to express metamodels underlying known MAS descriptors: MAS-ML (representing TAO) and Islander. As a result of that preliminary evaluation we noted an extension of our metamodel entailing new (dynamic) relationships amongst roles and between roles, as well as dynamic constructs such as Environment History to monitor message flows as described in [36]. This paper extends that evaluation to provide further evidence for the expressive power of our language constructed as a formalised synthesis of the implicit modelling approaches found in a number of existing agent-oriented methodologies. Based on the analysis in this paper, the initial version of the generic FAML metamodel was found to support most of the features in Gaia [8], Tropos [20], Adelfe’s [16], Prometheus [10], PASSI [11], MAS-CommonKASDS [21], MASE [22], RAP [23] MESSAGE [24] and Ingenias [25]. However, it was found to be deficient with respect to goal analysis, organizational structure, support for mobile agents, interaction protocol and system architecture (Table 4). We suggest adding support for goal analysis and organizational structure. However, for interaction protocol and system architecture, we view these as too methodology specific. They can be elaborated or constructed from our metamodel concepts. Support for mobile agents is the basis for future work. Beyond these extensions, we will undertake further validation against the underlying metamodels of a number of prominent methodologies, including Ingenias [25] and Tropos [20]. We also plan to further identify and exemplify its individual elements in the analysis of an actual P2P retrieval MAS application. Beyond the metamodel validation, the next step of our work is to create a complementary generic process metamodel and to situate the presented agent-oriented modelling language within a full method engineering framework. The modelling language
Towards Method Engineering for Multi-Agent Systems
265
will be stored in a repository as a collection of method fragments, which will be subsequently linked to other method fragments describing potential activities, tasks, techniques (i.e. process aspects), teams and roles (i.e. people aspects). Thus, a complete methodological framework will be provided, which will be able to support the generation of complete, custom-made agent-oriented methodologies using the tenets of method engineering.
Acknowledgement This work is supported by a GoldStar Award from University of New South Wales and the Australian Research Council.
References 1. Horlait, E.: Mobile Agents for Telecommunication Applications (Innovative Technology Series: Information Systems and Networks). Kogan Page, Portland (2004) 2. Guessoum, Z., Rejeb, L., Durand, R.: Using adaptive Multi-Agent Systems to Simulate Economic Models. In: AAMAS 2004. ACM, New York (2004) 3. Ferber, J., Drogoul, A.: Using Reactive Multi-Agent Systems in Simulation and Problem Solving. In: Avouris, L.G.N.M. (ed.) Distributed AI: Theory and Praxis. Kluwer, Brussels (1992) 4. Tidhar, G., et al.: Using Intelligent Agents in Military Simulation or Using Agents Intelligently. In: 11th Conference on Innovative Applications of Artificial Intelligence Papers (IAAI 1999), Orlando,Florida. MIT Press, Cambridge (1999) 5. Hunsberger, L., Grosz, B.J.: A combinatorial auction for collaborative planning. In: 4th International Conference on Multi-Agent Systems (ICMAS 2000) (2000) 6. Hogg, T., Williams, C.: Solving the Really Hard Problems with Cooperative Search. In: 11th National Conference on Artificial Intelligence, Washington, DC, USA. MIT Press, Cambridge (1993) 7. Edmonds, B., Bryson, J.: The Insufficiency of Formal Design Methods - the necessity of an experimental approach. In: AAMAS 2004. ACM, New York (2004) 8. Wooldridge, M., Jennings, N.R., Kinny, D.: The Gaia Methodology for Agent-Oriented Analysis and Design. In: Autonomous Agents and Multi-Agent Systems, The Netherlands. Kluwer Academic Publishers, Dordrecht (2000) 9. Bernon, C., Gleizes, M.-P., Peyruqueou, S., Picard, G.: ADELFE: A methodology for adaptive multi-agent systems engineering. In: Petta, P., Tolksdorf, R., Zambonelli, F. (eds.) ESAW 2002. LNCS, vol. 2577, pp. 156–169. Springer, Heidelberg (2003) 10. Padgham, L., Winikoff, M.: Developing Intelligent Agent Systems. A Practical Guide, vol. 1, p. 225. J. Wiley & Sons, Chichester (2004) 11. Cossentino, M., Potts, C.: A CASE tool supported methodology for the design of multi-agent systems. In: International Conference on Software Engineering Research and Practice (SERP 2002), Las Vegas, NV, USA (2002) 12. Bernon, C., Cossentino, M., Gleizes, M.-P., Turci, P., Zambonelli, F.: A study of some multi-agent meta-models. In: Odell, J.J., Giorgini, P., M¨uller, J.P. (eds.) AOSE 2004. LNCS, vol. 3382, pp. 62–77. Springer, Heidelberg (2005)
266
G. Low et al.
13. Beydoun, G., et al.: Developing and Evaluating a Generic Metamodel for MAS Work Products. In: Garcia, A., et al. (eds.) Software Engineering for Multi-Agent Systems IV: Research Issues and Practical Applications, pp. 126–142. Springer, Berlin (2006) 14. Brinkkemper, S.: Method Engineering: Engineering of Information Systems Development Methods and Tools. Information and Software Technology 38(4), 275–280 (1996) 15. Henderson-Sellers, B., Simons, A., Younessi, H.: The OPEN Toolbox of Techniques, Harlow (Essex), UK. The OPEN Series. Addison-Wesley Longman, Amsterdam (1998) 16. Bernon, C., Gleizes, M.-P., Peyruqueou, S., Picard, G.: ADELFE: A methodology for adaptive multi-agent systems engineering. In: Petta, P., Tolksdorf, R., Zambonelli, F. (eds.) ESAW 2002. LNCS, vol. 2577, pp. 156–169. Springer, Heidelberg (2003) 17. Esteva, M., Cruz, D.d.l., Sierra, C.: ISLANDER: an electronic institutions editor. In: International Conference on Autonomous Agents & Multiagent Systems (AAMAS 2002), Italy. ACM, New York (2002) 18. da Silva, V.T., Choren, R., de Lucena, C.J.P.: Using the MAS-ML to model a multi-agent system. In: Lucena, C., Garcia, A., Romanovsky, A., Castro, J., Alencar, P.S.C. (eds.) SELMAS 2003. LNCS, vol. 2940, pp. 129–148. Springer, Heidelberg (2004) 19. Tran, Q.N.N., Low, G.C.: Comparison of Ten Agent-Oriented Methodologies. In: Giorgini, B.H.-S.P. (ed.) Agent-Oriented Methodologies, pp. 341–367. Idea Group, Hershey (2005) 20. Giunchiglia, F., Mylopoulos, J., Perini, A.: The tropos software development methodology: Processes, models and diagrams. In: Giunchiglia, F., Odell, J.J., Weiss, G. (eds.) AOSE 2002. LNCS, vol. 2585, pp. 162–173. Springer, Heidelberg (2003) 21. Iglesias, C.A., Garijo, M.: The Agent-Oriented Methodology MAS- CommonKADS. In: Henderson-Sellers, B., Giorgini, P. (eds.) Agent-Oriented Methodologies, pp. 46–78. IDEA Group Publishing (2005) 22. DeLoach, S.A., Kumar, M.: Multi-Agent Systems Engineering: An Overview and Case Study. In: Henderson- Sellers, B., Giorgini, P. (eds.) Agent-Oriented Methodologies, pp. 236–276. IDEA Group Publishing (2005) 23. Firby, R.J.: Adaptive Execution in Dynamic Domains. Yale University, Yale (1989) 24. Garijo, F.J., Gomez-Sanz, J.J., Massonet, P.: The MESSAGE Methodology for AgentOriented Analysis and Design. In: Henderson-Sellers, B., Giorgini, P. (eds.) Agent-Oriented Methodologies, pp. 203–235. IDEA Group Publishing (2005) 25. Pavon, J., Gomez-Sanz, J., Fuentest, R.: The INGENIAS Methodology and Tools. In: Henderson-Sellers, B., Giorgini, P. (eds.) Agent-Oriented Methodologies, pp. 236–276. IDEA Group Publishing (2005) 26. Eydoun, G., et al.: Cooperative Modeling Evaluated. International Journal of Cooperative Information Systems 14(1), 45–71 (2005) 27. Tran, Q.-N.N., Low, G., Williams, M.-A.: A preliminary comparative feature analysis of multi-agent systems development methodologies. In: Bresciani, P., Giorgini, P., HendersonSellers, B., Low, G., Winikoff, M. (eds.) AOIS 2004. LNCS, vol. 3508, pp. 157–168. Springer, Heidelberg (2005) 28. Wood, B., et al.: A Guide to the Assessment of Software Development Methods, Software Engineering Institute, Carnegie Mellon University (1988) 29. Jayaratna, N.: Understanding and Evaluating Methodologies: NIMSAD, a Systematic Framework. McGraw-Hill, New York (1994)
Towards Method Engineering for Multi-Agent Systems
267
30. Olle, T.W., Sol, H.G., Tully, C.J.: Information Systems Design Methodologies - A Feature Analysis. Elsevier Science Publishers, Amsterdam (1983) 31. Inc., T.O.A. A Comparison of Object-Oriented Development methodologies (1995), http://www.toa.com/smnn?mcr.html 32. Shehory, O., Sturm, A.: Evaluation of modeling techniques for agent-based systems. In: Proceedings of the 5th International Conference on Autonomous Agents (2001) 33. O’Malley, S.A., DeLoach, S.A.: Determining when to use an agent-oriented software engineering paradigm. In: Wooldridge, M.J., Weiß, G., Ciancarini, P. (eds.) AOSE 2001. LNCS, vol. 2222. Springer, Heidelberg (2002) 34. Cernuzzi, L., Rossi, G.: On the Evaluation of Agent-Oriented Modeling Methods. In: Proceedings of the OOPSLA Workshop on Agent-Oriented Methodologies (2002) 35. Sabas, A., Badri, M., Delisle, S.: A Multidimentional Framework for the Evaluation of Multiagent System Methodologies. In: Proceedings of the 6th World Multiconference on Systemics, Cybernetics and Informatics (SCI 2002) (2002) 36. Beydoun, G., Debenham, J., Hoffmann, A.: Using Messaging Structure to Evolve Agents Roles. In: Barley, M., Kasabov, N. (eds.) Intelligent Agents and Multi-Agent Systems VII, pp. 18–30. Springer, Australia (2005)
Entrainment in Human-Agent Text Communication Ikuko Kanashiro1, Kazuki Kobayashi1, and Yasuhiko Kitamura2 1
Graduate School of Science and Technology, Kwansei Gakuin University 2-1 Gakuen, Sanda, Hyogo, 669-1337 Japan 2 School of Science and Technology, Kwansei Gakuin University, 2-1 Gakuen, Sanda, Hyogo, 669-1337 Japan {i_kanashiro,kby,ykitamura}@ksc.kwansei.ac.jp
Abstract. Non-verbal information such as utterance speed and switching pause create an impression of the speaker. If intelligent agents could handle such non-verbal information properly, the quality of interactions between agents and human users would improve. Entrainment is a phenomenon in which brainwave synchronization is established by means of periodic stimulus. It is reported that non-verbal information expressed by an individual entrains that expressed by another in voice communication. We have interest in how an agent can affect people though entraining the non-verbal information in text communication. Text is much easier for agents to handle than voice. Through experiments, we show that the utterance speed of an agent can entrain the duration of switching pauses used by human subjects. Keywords: Human-agent text communication, switching pause, utterance speed, entrainment.
establishes entrainment between two persons when they talk to each other smoothly. Nagaoka[2, 3] showed that switching pauses are entrained in voice communication. Entrainments between an artifact and people have also been observed. Komatsu et al. [5] report that utterance speed is entrained in voice communication between a person and a computer given the situation that the person and the computer interchangeably read sentences written in a script. If we can utilize entrainment properly, we can develop an artificial agent that can create desirable impressions by its non-verbal information [9]. Conventional works deal with entrainment in voice communication, but this paper deals with text communication. Text is much easier to be handled by agents than voice. Conventional text communication systems such as chat systems or messengers fail to well represent non-verbal information. Our solution is a chat system that changes that rate at which it displays individual characters in a message to alter the utterance speed. This paper shows how an agent can entrain non-verbal information such as utterance speed and switching pauses in text communication with humans. We believe that the result will lead to the development of agents that can create affective impressions by utilizing non-verbal information. Section 2 introduces a chat system that can control utterance speed. Section 3 shows how an agent can entrain the utterance speed and the switching pause of human users and we discuss the result in Section 4. We conclude this paper in Section 5 with our future work.
2 Chat System That Reflects Utterance Speed Users receive no impression of utterance speed in conventional chat systems because the systems display the whole message at once when received. We developed a chat system that can control the utterance speed. The utterance time per character (or just “utterance time”) T [msec/char] is defined as
T=
1 (T2 − T1 ) = C V
(1)
where T1 is the time when the sender starts typing his/her message, T2 is the time when he/she finishes typing it, and C is the number of characters in the message. The utterance speed V [char/msec] is defined as the inverse of T. Fig. 1 shows the components of the chat system that reflects the utterance speed of the message creator. Users A and B send messages to each other through client programs A and B, respectively. When user A inputs a message using client program A, the client sends it with the user's name and the utterance time to client program B through the server program. When client program B receives the message, it types the characters in the message one by one according to utterance time T. User B sends back his/her messages to user A and the message is processed likewise. This system allows each user to feel the utterance speed of his/her counterpart. The server program which mediates the client programs stores the history of chat data between users in a log file.
270
I. Kanashiro, K. Kobayashi, and Y. Kitamura
Fig. 1. Components of chat system that reflects utterance speed
3 Experiment We here examine how an agent can entrain the non-verbal information in text communication with subjects. The non-verbal information in voice communication has various modalities such as volume, pitch, and speed, but text communication is limited to just speed and switching pause. We examine how these two factors can establish entrainment. 3.1 Experiment System We replaced a client program of the chat system shown in Fig. 1 by an agent as shown in Fig. 2. The agent automatically sends messages to a subject. It doesn’t interpret the content of replies from the subject but merely uses them as cues for creating its next message. When the agent receives a cue, it reads the message (mi), utterance time (uti), and switching pause (spi) stored in the setting file. It then applies the switching pause and sends the message with the utterance time specified to the subject. The agent has no physical appearance like an avatar and all communication is through a chat interface as shown in Fig. 3. The subject receives a series of questions from the agent. We prepared only easy questions such as asking for the subject's name, age, preferences, and so on, in order to make him/her reply quickly. We didn’t use tough questions that would need time to reply because they may hinder smooth communication between the agent and the subject. As shown in Table 1, we prepared 52 questions and stored them with corresponding utterance time and switching pause in a setting file.
Entrainment in Human-Agent Text Communication
271
Fig. 2. Experimental system
Fig. 3. User interface Table 1. Example of questions from the agent Question (1) Hello. May I ask you some questions? What is your name? (2) What is your nickname? (3) How old are you? (4) …
Utterance Speed 500.00
Switching Pause 0.00
462.50 425.00 …
0.00 0.00 …
3.2 Procedure In this experiment, we examined how an agent can entrain non-verbal information expressed by subjects. Non-verbal information in text communication is limited to utterance speed and switching pause, so we examined how these factors establish entrainment. We assessed entrainment by measuring the following correlations.
272
I. Kanashiro, K. Kobayashi, and Y. Kitamura
(1) Correlation between utterance speed of the agent and that of the subjects (2) Correlation between utterance speed of the agent and switching pause of the subjects (3) Correlation between switching pause of the agent and that of the subjects (4) Correlation between switching pause of the agent and utterance speed of the subjects
40 university students participated in this experiment as subjects. We separated them into two groups; variable utterance speed group and variable switching pause group; each group had 20 subjects.
Each subject in the variable utterance speed group chatted with the agent whose utterance time varied from 50[msec/character] to 500[msec/character] in 13 steps of 37.5[msec/character] and with two peaks as shown in Fig. 4. The switching pause was fixed at 0[msec]. Each subject of the variable switching pause group chatted with the agent whose switching pause varied from 0[msec] to 9600[msec] with 13 levels separated by 800[msec] with two peaks as shown in Fig. 5. The utterance time was fixed at 0[msec/character] in this group so each message was displayed in its entirety. 3.3 Results We plot variations between the utterance speed of the agent and that of subjects in Fig. 6(a). The vertical axis represents the utterance time of the agent and the horizontal axis represents that of the subjects. Multiple-regression was used to determine the variations between the utterance speed of the agent and that of a subject [12]. We used the utterance speed of subjects as outcome variable, that of the agent and subjects as predictor variable. Subject was treated as categorical factor using dummy variable with 19 degrees of freedom. The p value from t test for the regression slope of the utterance speed of the agent was used to determine the probability of the analysis. The magnitude of correlation coefficient between the utterance speed of the agent and that of subjects was calculated as square root of (sum of squares for the utterance speed of the agent) / (sum of squares for the utterance speed of the agent + residual sum of squares). The sign of the correlation coefficient was given by that of the regression coefficient for the utterance speed of the agent. The correlation with correlation coefficient of 0.049 (p=0.111) was found, but this correlation is not significant. Fig. 6(b) shows variations between the utterance time of the agent and the switching pause of the subjects. The correlation coefficient is 0.661 (p<0.01), so the correlation is significant and relatively strong. Fig. 6(d) shows variations between the switching pause of the agent and that of the subjects. The correlation coefficient is -0.05 (p=0.109), so the correlation is not significant. Fig. 6(c) shows variations between the switching pause of the agent and the utterance time of the subjects. The correlation coefficient is 0.071 (p<0.05), so this correlation is significant but weak. Table 2 summarizes the experiment's results. The utterance speed of the agent entrains the switching pause of the subjects but the other combinations do not yield any significant entrainment. Table 2. Summary of experiment's results
utterance speed Agent switching pause
Subjects switching pause relatively strong correlation no significant weak correlation correlation utterance speed no significant correlation
I. Kanashiro, K. Kobayashi, and Y. Kitamura
Subjects’ switching pause [msec]
274
15000 10000 5000 0 0
100
200
300
400
500
Agent’s utterance time [msec/character]
4000
(b) Correlation between the utterance time of the agent and the switching pause of the subjects. Subjects’ switching pause [msec]
Subjects’ utterance time [msec/character]
(a) Correlation between the utterance time of the agent and that of the subjects.
3000 2000 1000 0 0
2000
4000
6000
8000 10000
Agent’s switching pause [msec]
(c) Correlation between the switching pause of the agent and the utterance time of the subjects.
15000 10000 5000 0 -5000 -10000 0
2000 4000 6000 8000 10000
Agent’s switching pause [msec]
(d) Correlation between the switching pause of the agent and that of the subjects.
Fig. 6. Correlations of the utterance time and the switching pause between the agent and the subjects
4 Discussion 4.1 Typing Ability of Subjects The above results show there is a relatively strong correlation between the utterance time of the agent and the switching pause of the subjects, but there is no significant correlation between the utterance time of the agent and that of the subjects. We discuss here the failure of the utterance time to establish entrainment. Table 3 shows the correlations between utterance time of the agent and that of each subject. It indicates that 5 of the 20 subjects showed significant correlations (p<0.05), though the correlation was weak. Fig. 7 shows the correlation between the utterance time of the agent and that of the subject (No.5) who showed the strongest correlation.
Entrainment in Human-Agent Text Communication
275
Table 3. Correlation between the utterance time of the agent and that of each subject
Agent’s utterance time [msec/character] Fig. 7. Correlation between the utterance time of the agent and that of subject No.5 whose correlation is strongest in the group
The reason why most subjects were not entrained by the agent may be due to their limited typing ability. Although the subjects may have been entrained with regard to utterance speed, they may not have been able to express it because of their limited typing ability. In future work, we would like to perform this experiment again using some approach to offset the subjects' typing skills. 4.2 Switching Pause of the Agent The experiment showed that the switching pause of the agent failed to establish entrainment. The reason may reflect how the subjects interpreted the switching pause of the agent. After this experiment, some subjects said they thought the chat system or the
276
I. Kanashiro, K. Kobayashi, and Y. Kitamura
network was out of order. The subjects thought that the change in switching pause was merely due to a system problem but not due to the agent. 4.3 Affective Influence on Subjects Changes in the agent’s utterance speed influenced the subjects. In the experiment, the utterance speed of the agent entrained the switching pause of the subjects. The increased utterance speed of the agent goaded the subjects to replying more quickly. In future work, we need to examine the affective influence of the agent’s utterance speed on the subjects in detail, for example by using a questionnaire survey or by measuring biometrics such as sweating or cardiac rate.
5 Summary Communication between people involves not only verbal information but also non-verbal information, and the non-verbal information also creates an impression of the speaker. We find that entrainment can be established by non-verbal information. Conventional studies on entrainment consider voice communication between two people or between a person and an agent, but we deal with entrainment in text communication. In this paper, we examined how the two factors of non-verbal information, utterance speed and switching pause, can entrain the responses of human subjects. For this experiment, we developed a chat system which uses typing speed to be used as an indicator of utterance speed. We conducted an experiment to show that the utterance speed of an agent entrains the switching pause of subjects. We could not observe any other significant entrainment among the other combinations of utterance speed and switching pause. We can conclude that the utterance speed is effective in establishing entrainment, and that switching pause is entrained by the utterance speed. In future work, we will examine the affective influence of utterance speed on people. We expect this work is lead to that development of interface agents that affectively interact with human users through the use of non-verbal information in text communication.
References 1. Nagaoka, C., Draguna, M., Komori, M., Nakamura, T.: The Influence of Switching Pauses on Interpersonal Perception in Dialogues. In: Proc. of Human Interface 2002, pp. 171–174 (2002) (in Japanese) 2. Nagaoka, C., Draguna, M., Komori, M., Kawase, S., Nakamura, T.: The Mutual Influence of the Speakers’ Switching Pauses in Dialogues. In: Proc. of Human Interface 2001, pp. 221–224 (2001) (in Japanese) 3. Nagaoka, C., Komori, M., Nakamura, T.: The interspeaker influence of the switching pauses in dialogue. Japanese Journal of Ergonomics 38(6), 316–323 (2002) 4. Watanabe, T., Okubo, M.: Physiological Analysis of Entrainment in Face-to-Face Communication. In: Proceedings of the Seventh International Conference on Human-Computer Interaction, pp. 411–414 (1997)
Entrainment in Human-Agent Text Communication
277
5. Komatsu, T., Morikawa, K.: Entrainment in the Rate of Utterance in Speech Dialogs Between Users and an Auto Response System. Journal of Universal Computer Science 13(2), 186–198 (2007) 6. Fogg, B.J.: Persuasive Technology: Using Computers to Change What We Think and Do. Morgan Kaufmann, San Francisco (2003) 7. Vargas, M.F., Words, L.T.: An Introduction to Nonverbal Communication. Lowa State University Press (1987) 8. Patterson, M.L.: Nonverbal Behavior: A Functional Perspective. Springer, Heidelberg (1983) 9. Picard, R.: Affective Computing. MIT Press, Cambridge (1997) 10. Cassell, J., Sullivan, J., Prevost, S., Churchill, E.: Embodied Conversational Agents. MIT Press, Cambridge (2000) 11. Prendinger, H., Ishizuka, M. (eds.): Life-Like Characters. Springer, Heidelberg (2004) 12. Bland, J.M., Altman, D.G.: Calculating correlation coefficients with repeated observations: Part 1 – correlation within subjects. Brit. Med. J. 310, 446 (1995) 13. Reeves, B., Nass, C.: The Media Equation. CSLI Publications (1998)
A Driver Modeling Methodology Using Hypothetical Reasoning for Multiagent Traffic Simulation Yusuke Tanaka, Yuu Nakajima, Hiromitsu Hattori, and Toru Ishida Department of Social Informatics, Kyoto University Yoshida Honmachi, Sakyo-ku, Kyoto, 606-8501 Japan {ytanaka,nkjm}@ai.soc.i.kyoto-u.ac.jp, {hatto,ishida}@i.kyoto-u.ac.jp
Abstract. We propose how to acquire driver’s individual operation models using the three-dimensional driving simulator in order to implement distinct personalities on each agent. In this paper, operation models are defined as sets of prioritized operation rules, each of which consists of the world as observed by a driver and his/her next operation according to the observation. Each driver might have different set of rules and their priorities. We apply a method to acquire individual operation models using hypothetical reasoning. Because of the method, we are able to obtain operation models which can explain driver’s operation during driving simulation. We show some operation models acquired from aged/young human drivers, and then clarify the proposed method can catch each driver’s characteristics.
1
Introduction
Multiagent-based traffic simulation has been considered as one of the promising approach to analyze traffic flow [1,2]. In the multiagent traffic simulation, each human driver is modeled as an agent, which has common properties for driving. In the real world, there is a variety of drivers, such as aged drivers, novice drivers, and so on. In order to incorporate such variety of driver’s properties into the traffic simulation, it is required to develop techniques to construct human driver’s operation model [3,4,5]. We propose how to acquire driver’s individual operation models using the three-dimensional driving simulator in order to implement distinct personalities on each agent. In this paper, we define an operation model as sets of prioritized operation rules, each of which consists of the world as observed by a driver and his/her next operation according to the observation. Each driver might have different set of rules and their priorities. In this paper, we address the following two research issues: 1) Forming individual operation models from driving simulation log In order to acquire diverse driver models, we extract individual operation models from driving simulation log data. A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 278–287, 2009. c Springer-Verlag Berlin Heidelberg 2009
A Driver Modeling Methodology Using Hypothetical Reasoning
279
2) Efficient acquisition of individual operation models In order to form individual operation models, time-consuming tasks are required, such as recruitment of examinees, analysis of simulation log data. Therefore, we try to enable formation of operation models from little log data and operation rules, which are extracted through the interview with the small number of examinees. For achieving these two issues, we apply a method using hypothetical reasoning for modeling driver’s operation [6]. To put it concretely, driver’s operation models are acquired according to the following process; 1) acquisition of examinees’ behavior, 2) formal description of the observations in predicate logic, 3) extraction of operation rules through the interviews, and 4) construction of operation models from collected behavior and operation rules based on hypothetical reasoning. From the viewpoint of traffic engineering, it is sometimes unclear what operation rules are used by an individual driver when he/she is driving a vehicle. In contrast, we can acquire operation models with only operation rules by using a method based on hypothetical reasoning. The remainder of the paper is organized as follows. First, we describe the modeling process using hypothetical reasoning with log data resulting from driving simulation. Second, we define technical terms used in this paper, and then formalize our target problem by using the defined terms for hypothetical reasoning. Then, we apply a learning method using hypothetical reasoning to driving simulations for acquisition of operation models. Finally, concluding remarks are given in the final section and we also discuss possible future directions.
2
Modeling Process Using Driving Simulation Log
We explain an agent modeling process using hypothetical reasoning in our research. In the process, first, we conduct driving simulation on the three-dimensional driving simulator. Then, we acquire an examinee’s operation model, a set of operation rules, by explaining the examinee’s behavior observed in one or more driving simulations. However, the obtained operation rules may exhibit some incompatibility. Therefore, we hypothesize whether each operation rule is employed by the target examinee, and choose the assumptions that pass hypothetical reasoning, which offers the consistent selection of hypotheses. The result of hypothetical reasoning is a set of compatible operation rules employed by the target examinee. Some log data are not explained using operation rules, and then such data is removed. Details of an agent modeling process is as follows (also shown in Figure 1): 1) Accumulation of driving log: Conduct driving simulations where each examinee drives a virtual car on the driving simulator. 2) Log data cleaning: Eliminate operation log data which cannot be explained by operation rules. 3) Formal description of the observations: Collect an observation from log data in driving simulations and describe the observations in predicate
280
Y. Tanaka et al.
Fig. 1. Modeling Process
logic. In this paper, “observation” means the world (environment) which is observed by an examinee. 4) Interview with examinees: Interview with examinees using screen capture during driving simulation and kinds of charts. 5) Collecting operation rules: Collect operation rules constituting domain knowledge from the result of interview. In this paper, each operation rule represents what examinees observed and how they operated during simulations. 6) Learning models using hypothetical reasoning: Acquire candidate operation models using hypothetical reasoning with the domain knowledge and the observations. As we mentioned above, operation models are defined as sets of prioritized operation rules. In this paper, we focus on how to acquire operation models using hypothetical reasoning from log data resulting from driving simulations.
3
Formalization of the Problem in Driving Simulation
In this section, in order to apply hypothetical reasoning to the acquisition of operation models, we formally define operation rules, observations, and operation models. 3.1
Operation Rules
We describe an operation rule as a condition-action rule. The condition part of a rule describes an examinee’s operation and the action part describes situation observed by an examinee. If and only if all conditions are satisfied, the action part of the rule could be executed. Example 1 shows a description of operation rules. Example 1: Description of operation rules rule1 : if Curve(x),InSight(x,self) then LoosenAccel(self) rule2 : if U phill(x),InSight(x,self) then StrengthenAccel(self) rule3 : if M oreT hanDesiredSpeed(self) then LoosenAccel(self)
A Driver Modeling Methodology Using Hypothetical Reasoning
281
rule1 : if an examinee(self) sees(InSight) a curve x(Curve), he/she releases the accelerator(LoosenAccel). rule2 : if an examinee sees an uphill road x(U phill), he/she depresses the accelerator(StrengthenAccel). rule3 : if the speed of examinee’s vehicle is over his/her desired speed(M oreT hanDesiredSpeed), he/she releases the accelerator. 3.2
Observation
The observation, which is included in log data, is described in predicate logic according to the time line. We use road shape, driving speed, and acceleration pedal operation as observations. An example of the description of observation is as follows: Example 2: Description of observation Curve(Curve2 ) ∧ InSight(Curve2 , self) ∧ U phill(U phill3) ∧ On(U phill3 , self) ∧ Accelerate(self) ∧ M oreT hanDesiredSpeed(self) ∧ StrengthenAccel(self) ⇒ Do(LoosenAccel(self)) This observation means that an examinee releases the accelerator when the examinee sees Curve2 , his/her vehicle is running on U phill3(On), he/she is increasing speed(Accelerate), the speed of vehicle is over his/her desired speed, and he/she depresses the accelerator. In this example, Do is a predicate meaning that he/she initiates an operation. 3.3
Operation Model
An operation model consists of sets of prioritized operation rules which are actually used by an examinee. If the condition part of an operation rule is satisfied, the rule is available to use. If there are many usable rules, the one which has the highest priority is chosen. Example 3 shows a description of an operation model. Example 3: Description of operation model (rule1 , rule2 , rule3 , rule2 rule1 ) The meaning of this example is as follows. An examinee uses rule1 , rule2 , and rule3 . The priority of rule1 is higher than that of rule2 , and thus when both of the condition part of rule1 and rule2 are satisfied at the same time, he/she uses rule1.
4
Detail of Driver Modeling
In this section, we are showing how to apply hypothetical reasoning to the acquisition of examinee’s operation models in driving simulation.
282
Y. Tanaka et al.
Fig. 2. Log data with the result of the Interviews
4.1
Driving Simulation
In our research, we try to acquire log data from aged examinees and young examinees, and then investigate their decision making process to determine their operation. During driving simulation, each examinee drives 11km road including diverse road alignments. In order to focus on the relationship between each examinee’s operations and road conditions, we choose solo driving scenario. Through the investigation, we analyze the relationship between each examinee’s operations and road conditions. Figure 21 shows an example of log data with answers at post simulation interview. The horizontal axis denotes the elapsed time, and the vertical axis denotes speed (km/h), acceleration (m/ss), the rate of acceleration/brake(%). 4.2
Collecting and Cleaning-Up Log Data
In this research, we use the following data to construct operation models. 1) 2) 3) 4)
Time(s): The elapsed time from the beginning of driving simulation. Mileage(km): The mileage from the start point. Speed(km/h): The speed of examinee’s vehicle Rate of Accel./Brake(%): The rate of acceleration/braking, i.e., accelerator/brake pedal position. When an examinee does not step on the pedal, the rate is 0%, and he/she fully steps on the pedal, the rate is 100%.
During driving simulation, there is a possibility of examinee’s unintentional/ meaningless operation. For example, an examinee sometimes increases the speed without any explicit reasons. Currently, we only have a method to acquire models from examinee’s intentional operation. This is because an operation rule consists 1
The log data is offered by Iida Lab., Osaka University.
A Driver Modeling Methodology Using Hypothetical Reasoning
283
of intentional operation and surrounding environment, so that it is impossible to explain an unintentional operation by using operation rules. Therefore, we eliminate unintentional operation log before the step for acquiring operation models. Currently, we take the following two policies for the elimination. 1) Elimination of unintentional operations In general, it is practically difficult to execute intentional operations in short time. Thus, we eliminate such operations from log data. We empirically eliminate operations which are subsequently executed within 2 seconds (shown in Figure 3). 2) Elimination of operations unrelated to surrounding environment An examinee sometimes continuously executes operations until he/she can recognize the change of surrounding environment. For example, when an examinee wants to reduce the speed, he/she steps on the brake several times. This is because it takes time to get to the desired speed. In such case, there are many moves in log data due to the continuous operations. Therefore, we use only the first operation and eliminate subsequent operations. Empirically, we eliminate operations subsequently executed within 3.5 seconds (also shown in Figure 3).
4.3
Description of Observation
We describe observations, extracted from log data, based on predicate logic. In this paper, we pick up observations which could affect examinees’ operations, and define the following predicates. Straight(x), Curve(x), U phill(x), Downhill(x): x is straight/curve/ uphill/downhill On(x, y): y is running on x F inish(x, y): y is running at the end of curve x InSight(x, y): y can see x Sharp(x), Slow(x): x(e.g. curve, slope) is sharp or slow M oreT hanDesiredSpeed(x), LessT hanDesiredSpeed(x): the speed of vehicle x is over the desired speed M oreT hanCurveSpeed(x, y): the speed of vehicle y is too high to safely drive in curve x Accelerate(x), Decelerate(x): the vehicle x is accelerating/decelerating KeepAccel(x): x keeps the rate of acceleration StrengthenAccel(x): x is increasing the rate of acceleration LoosenAccel(x): x is decreasing the rate of acceleration ReleaseAccel(x): x is releasing the accelerator Brake(x): x is stepping on the brake
284
Y. Tanaka et al.
Fig. 3. Types of Elimination
4.4
Interview
We interview an examinee with a chart which shows speed/acceleration/rate of accelerator and brake. Additionally, we use screen capture of an examinee’s screen in order to help him/her to remember his/her behavior during simulation. In the interview, we ask about the following four points: 1) Cause of gaze/Object of gaze: The reason why an examinee gazes something or the thing what an examinee gazes. 2) Recognized object: The thing what an examinee recognizes 3) Evaluation: The feeling that an examinee has 4) Reason for the decision: The reason of examinee’s decision making In this paper, we have an interview on some points where the speed/acceleration is drastically changed (5km/h, 5m/ss) or his/her rate of acceleration is drastically changed (10%). Examples of answers to above four points are shown in Figure 2. 4.5
Extraction of Operation Rules
We analyze the result of interview and log data, then we generate an operation rule which can denote an examinee’s generic operation. We extract the condition part of a rule by analyzing the result of interview (cause of gaze, recognized object). We extract the action part of a rule by analyzing examinee’s operations for the condition. In this paper, we extract operation rules from aged and young examinees. As a result, the following operation rules are extracted. rule1: if an examinee sees a curve ahead, he/she releases the accelerator rule2: if an examinee sees a curve ahead and the speed of vehicle is too high to drive the curve, he/she releases the accelerator rule3: if an examinee sees a sharp curve ahead and the speed of vehicle is too high to drive the curve, he/she releases the accelerator rule4: if an examinee is driving at the end of a curve, he/she depresses the accelerator rule5: if an examinee sees an uphill ahead, he/she depresses the accelerator rule6: if an examinee sees a downhill ahead, he/she releases the accelerator rule7: if an examinee is driving on an uphill, he/she depresses the accelerator rule8: if an examinee is driving on a downhill, he/she releases the accelerator
A Driver Modeling Methodology Using Hypothetical Reasoning
285
rule9: if the speed is over the desired speed, an examinee releases the accelerator rule10: if an examinee is driving on a straight road and the speed is under the desired speed, he/she depresses the accelerator rule11: if an examinee depresses the accelerator and then the vehicle is accelerating, he/she keeps the rate of the accelerator rule12: if an examinee releases the accelerator and then the vehicle is decelerating, he/she keeps the rate of the accelerator. 4.6
Operation Models
In this section, we show operation models acquired by our proposed method. Example 1: Operation model acquired from the aged examinee A04 ({rule7 , rule8 , rule11 , rule12 }, rule7 = rule8 rule11 , rule7 = rule8 rule12 ) This model means that the aged examinee A04 used four operation rules (rule7 , rule8 , rule11 , rule12 ). Figure 4(a) shows the order of priorities of these rules. In this figure, the direction of an arrow describes the order of priority between rules. An rule with an arrow-head has the higher priority. For example, the priority of rule11 is higher than that of rule7 and rule8 . rule7 = rule8 means that rule7 rule8 and rule8 rule7 are approved. In the figure, an arrow in both direction shows that the priorities of both rules are same. Example 2: Operation model acquired from the aged examinee A20 ({rule5 , rule7 , rule9 , rule11 , rule12 }, rule5 = rule7 = rule9 rule11 , rule5 =rule7 =rule9 rule12 ) This model means that the aged examinee A20 used five operation rules (rule5 , rule7 , rule9 , rule11 , rule12 ). Figure 4(b) shows the order of priorities of rules. Because the order of priorities of three rules (rule5 , rule7 , rule9 ) are same, this model represents that the examinee depressed or released the pedal when he/she saw an uphill and the speed of his/her vehicle was over the desired speed. Therefore, it would appear that he/she could not always sense an uphill and the driving speed was over his/her desired speed. Example 3: Operation model acquired from the young examinee Y01 ({rule7 , rule11 }) This simple model (Figure 4(c)) means that the young examinee Y01 drove based on only two operation rules (rule7 , rule11 ). Because the condition part of these rules could not be satisfied at the same time, the order of priorities of these rules are not fixed. Example 4: Operation model acquired from the young examinee Y08 ({rule7 , rule8 , rule9 , rule11 , rule12 }, rule7 rule8 =rule9 , rule7 rule12 ) This model (Figure 4(d)) means that the young examinee Y08 used five operation rules (rule7 , rule8 , rule9 , rule11 , rule12 ). Figure 4 shows the order of priorities of these rules. Because the priority of rule9 is higher than that of rule7 , it
286
Y. Tanaka et al.
Fig. 4. The order of priorities of operation rules
would appear that the examinee wanted to keep his/her desired speed instead of keeping up the speed on an uphill. In this research, operation models of a young examinee usually consists of less numbers of operation rules. On the other hand, the models of an aged examinee consists of more numbers of rules. That is to say, an young examinee tends to drive a vehicle based on simple driving model and an aged one tends to drive a vehicle based on relatively complex model.
5
Conclusion
For conducting multiagent-based traffic simulation, one of the central issue is how to model an agent including diverse characteristics of a human driver. For achieving this issue, we proposed a method to acquire operation models using hypothetical reasoning, and then showed acquired operation models from aged/young human drivers. The main contributions of this paper are the following two points: 1) A method to acquire operation models from log data in driving simulation We have proposed a method to acquire operation models of an individual examinee using log data in driving simulation. In the proposed method, first the interview is conducted to catch the reasons of behaviors during driving simulation, and operation rules are described using predefined predicates. Then, we use hypothetical reasoning to acquire characteristic operation models. The acquired model is sets of prioritized operation rules.
A Driver Modeling Methodology Using Hypothetical Reasoning
287
2) Reduction of modeling cost Actually, having the interview and log data analysis are time-consuming task. However, in our method, firstly operation rules are acquired from some examinees. Then, all acquired rules are usable for explaining any examinee’s log data. We do not need to interview all examinees and analyze all log data, so thus we could cut out troublesome modeling cost. In the proposed method, it is possible to acquire promising operation models if an examinee takes simple “intentional” operations. That is to say, when we use straightforward log data, the proposed modeling process using hypothetical reasoning could work well. If there is noise data, i.e. “unintentional” operations, it becomes difficult to accurately acquire operation model. Our possible future work includes: improving noise-canceling method in order to acquire more sophisticated operation model; incorporating more information (condition) into operation rules; implementation of the models, operating on traffic simulation, based on the acquired operation models; analyzing the validity of the models for practical traffic simulation by making evaluation experiments.
Acknowledgment The authors would like to thank K. Iida for making this work possible. This research was partially supported by a Grant-in-Aid for Scientific Research (A) (18200009, 2006-2008) from Japan Society for the Promotion of Science (JSPS).
References 1. Paruchuri, P., Pullalarevu, A.R., Karlapalem, K.: Multi agent simulation of unorganized traffic. In: Proceedings of The First International Joint Conference on Autonomous Agents and Multiagent Systems (AAMAS 2002), pp. 176–183 (2002) 2. Balmer, M., Cetin, N., Nagel, K., Raney, B.: Towards truly agent-based traffic and mobility simulations. In: The Third International Joint Conference on Autonomous Agents and Multiagent Systems (AAMAS 2004), pp. 60–67 (2004) 3. Espie, S.: Vehicle-driven simulator versus traffic-driven simulator: the inrets approach. In: Driving Simulation Conference (DSC 1999), pp. 367–376 (1999) 4. Bonabeau, E.: Agent-based modeling: Methods and techniques for simulating human systems. Proceedings of the National Academy of Sciences of the United States of America 99(10), 7280–7287 (2002) 5. Fernlund, H., Gonzalez, A.J., Gerber, W.J., DeMara, R.F., Georgiopoulos, M.: Learning tactical human behavior through observation of human actor. IEEE Transactions on Systems, Man and Cybernetics (2005) 6. Murakami, Y., Sugimoto, Y., Ishida, T.: Modeling human behavior for virtual training systems. In: The 20th National Conference on Artificial Intelligence (AAAI 2005), pp. 127–132 (2005)
Analysis of Pedestrian Navigation Using Cellular Phones Yuu Nakajima1 , Takatoshi Oishi1 , Toru Ishida1 , and Daisuke Morikawa2 1
Department of Social Informatics, Kyoto University Yoshida-Honmachi, Sakyo-ku, Kyoto, 606-8501, Japan {nkjm,oishi,ishida}@ai.soc.i.kyoto-u.ac.jp 2 KDDI R&D Laboratories Inc. 2-1-15 Ohara, Fujimino, Saitama, 356-8502, Japan [email protected]
Abstract. Navigation services for pedestrians are spreading in recent years. Our approach to provide personal navigation is to build a multiagent system that assigns one guiding agent to each human. This paper attempts to demonstrate a design implication of the guiding agent. In the navigation experiment where a pedestrian using a map on a GPScapable cellular phone was guided by a distant navigator, we observed the communication between them by conversation analysis. The result suggests that information required by a pedestrian were the current location, the current direction and a proper route toward a destination. The communications between a pedestrian and a navigator were based on a navigation map or a movement history. When a pedestrian did not understand the map adequately, navigation sometimes failed due to the lack of communication basis.
1
Introduction
Traveling in an unfamiliar city is a daily task for ordinary people. For instance, they look for meeting spots or shops in unfamiliar cities. These days, more and more pedestrians use cellular phones as information sources for route guidance. Pedestrians use cellular phones in two ways as information sources. One is displaying a map showing the current location, and the other is consulting with a distant navigator via voice conversations. Due to the popularity of and improvement in sensor devices and network devices, environments that support ubiquitous computing are spreading. In such environments, it is possible to provide personal navigation that suits the properties, the location and the context of each user [1]. We built evacuation navigation system based on multi-agent server, which assigns one guiding agent to each human. In this system, an agent can provide a personalized navigation map considering the current location and the surrounding environment [2][3]. Some people may not be able to reach their destination even if they use such navigation systems. People who are not good at reading maps should ask others for help. This paper assumes that pedestrians take part in evacuation drill with A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 288–297, 2009. c Springer-Verlag Berlin Heidelberg 2009
Analysis of Pedestrian Navigation Using Cellular Phones
289
Fig. 1. Large-scale evacuation guide system with guiding agents
the evacuation navigation system. To demonstrate requirements of the guiding agent, we address the following two issues. Analysis of information required by pedestrian. To examine a design implication of the guiding agent, it is necessary to investigate the information requirements of pedestrians when they use a navigation map. Analysis of communication between pedestrian and navigator. A pedestrian cannot be always guided into a proper route by a distant navigator. Investigation of failure cases is needed in order to examine the limitation of remote navigation. We analyze the communication between the pedestrian and the navigator using conversation analysis1 .
2
Large-Scale Evacuation Guide System
We produced a large-scale evacuation guide system based on large scale agent platform and GPS-capable cellular phones. Fig. 1 depicts system architecture of pedestrian navigation system [3]. In a navigation system which uses ubiquitous information infrastructure on a city, the system can acquire information of each individual user in real time. However, quantity of the information becomes enormous. There occurs a problem that a human who control system cannot handle all the information. Our approach is that a human gives rough navigation to agents and the agents give precise navigation to each person. We aim at realizing a mega scale navigation system using GPS-capable cellular phones. The control interface was implemented based on transcendent communication architecture. Transcendent communication is proposed as the method for navigation in public spaces [5]. In transcendent communication, the distribution of 1
Conversation analysis is a methodology for studying social interaction. It was principally developed by Harvey Sacks and Emanuel Schegloff [4].
290
Y. Nakajima et al.
Fig. 2. User interface of navigation system Fig. 3. Navigation map on cellular phone
evacuees in the real space is reproduced on the virtual space as human figures that mirror the positions of evacuees; the positions of the subjects are acquired by sensors. The system commander assigns evacuation destinations and evacuation directions through the control interface shown in Fig. 2. The commander issues high level instructions to the guiding agents using a map. The guiding agents that assigned to evacuees on a one-to-one basis provide individual navigation maps via GPS-capable cellular phones. An agent is instructed on a direction of evacuation by the control center. The agent retrieves shelters around the user, and selects a destination according to the ordered direction and distance between the user and each shelter. The agent also obtains the surrounding environmental data from a database. Then, the agent sends their users personalized navigation map showing a destination, the direction, impassible point, the current location and the movement history as indicated on Fig. 3.
3 3.1
Experiment of Pedestrian Navigation Overveiw of Experiment
We conducted two navigation experiments in which pedestrians were guided by distant navigators using the large-scale evacuation navigation system. These experiments have two purposes. One is to investigate the information required by pedestrians, and the other is to analyze communication between the pedestrian and the navigator. May et al. already examined information requirements in an experiment of navigation based on turn-by-turn strategy [6]. However the human subjects did not take along a map showing the current location in the research. We analyzed the information requirements with think aloud method [7] which was to observe what human subjects behave and think. The method has actually been applied to experiments in navigation systems. In the method, an experimenter instructs human subjects to think aloud while performing tasks. The experimenter observes the human subjects’ behavior and thinking at the same
Analysis of Pedestrian Navigation Using Cellular Phones
291
Fig. 4. Experiment 1: observation of conversation between pedestrians
Fig. 5. Experiment 2: observation of conversation between pedestrian and navigator
time. The experimenter can combine human subjects’ behavior with their feelings about the system which they have in mind. In these navigation experiments, we observed two kinds of conversations. Observation of conversation between pedestrians. The pedestrians used navigation system which provided a map as Fig. 3. In the first experiment, we let human subjects to use system in pairs and instructed them to talk to each other (Fig. 4). We refer to this experiment as experiment 1. Observation of conversation between pedestrian and navigator. In the second experiment, pedestrians used the navigation system and consulted with distant navigators via voice conversations (Fig. 5). We instructed pedestrians to talk to navigators anytime they have a question. We call this experiment as experiment 2.
292
Y. Nakajima et al.
In order to analyze the relativity between the conversation and the behavior, we needed to gather actions, speeches and eye sights of human subjects. Following three data were collected; 1) recordings of the conversations between pedestrians and pedestrian or pedestrian and navigator, 2) video pictures of the behavior of pedestrians, 3) video pictures of the sight pedestrians have seen. In experiment 1, a cameraman followed a pair of pedestrians and shot their behaviors with video camera. We recorded speeches of the pedestrians with attached microphones and transmitted them to the video camera by Bluetooth audio transmitters. In experiment 2, a cameraman followed a pedestrian and shot his behavior with video camera as in the case of experiment 1. Maps provided to pedestrians and navigation screens were recorded by DV video recorder. Conversations between the pedestrian and the navigator were also recorded with call-recording microphone. After the experiments, the graphic data was synchronized with the voice data. 3.2
Scenario of Experiment
We assumed the scenario that a huge earthquake struck around Kyoto University and dwellers escaped to shelters using navigation system through GPS-capable cellular phones. We chose a evacuation drill as scenario for the reason that pedestrians must choose their route carefully and more depended on the system than usual. We preliminarily set disaster points and subsequently added secondary disaster points during the experiment in order to make route selection difficult and to make pedestrians behave carefully. The direct distance from a start to a destination in the experiment was about 1.5 km. The number of human subjects performed as pedestrians was four. All of them were students and unfamiliar with the testing area. Additionally, they had no experience using the evacuation navigation system. In experiment 1, the pedestrians performed the task in pairs so that we gathered two groups of data. In experiment 2, the pedestrians performed the task alone so that we gathered four groups of data. 3.3
Task and Instruction
In experiment 1, pairs of pedestrians left start points and headed for destination points. They consulted a map showing the current location on a cellular phone and talked to each other on their travel. In experiment 2, pedestrians had two cellular phones so that they could consult with a distant navigator in parallel with reading a map. When they have any questions or any troubles, they could consult the navigator to resolve it. In experiment 2, the task of a navigator was to answer questions from a pedestrian. The navigator was expected to make the maximum efforts to answer any questions from a pedestrian. The navigator should use information on the navigation system sufficiently. An information screen for a navigator is shown in Fig. 2. A navigator could view environmental information about a local area, the current locations of pedestrians and a map provided to a pedestrian. Considering
Analysis of Pedestrian Navigation Using Cellular Phones
293
Table 1. Conversation about current location
A: B: A: B:
((The corner to turn is)) here? What do you think about that? Further, isn’t it further? Because it is still What? But it ((= map)) has not reloaded yet. I see.
those information, the navigator guided the pedestrian via voice conversation. The navigator answered the questions consulting a screen of navigation system. Experimenters who were familiar with the experimental area acted as navigators. Additionally they had practiced guiding pedestrians beforehand. They guided pedestrians one-on-one in the experiment. The purpose of the experiment is to analyze what kind of information pedestrians required. Therefore a navigator must not speak to a pedestrian voluntarily. Pedestrians ware expected to be aware of the necessity to select routes carefully and to be encouraged to voice what they are thinking in these experiments. To ensure it, we gave the instructions below to the human subjects in the experiments; “This is an evacuation drill. Pedestrians are expected to act carefully and to commit themselves to reach shelters in safety”, “Disaster points are impassable so that pedestrians must avoid them”.
4 4.1
Result and Analysis Information Required by Pedestrians
Conversations about questions, confirmations and trouble were extracted from pedestrians’ speeches and transcribed. These transcripts were categorized according to pedestrian’s intentions or pedestrian’s demands. The result shows that the pedestrians required information about the current location, the current direction and a proper route to a destination. The information was used to confirm and trust a route as well as decide it. The information required by pedestrians are categorized into the following three types. Current location. In Table 1 case, pedestrians could not have the confidence in correspondence between a map and the real world. Such cases were frequently observed when pedestrians could not recognize the corner that they had decided to turn as reading a map in advance. Current direction. In Table 2 case, pedestrians could not understand a map properly because of losing sense of direction. Sense of direction was important for understanding maps. Proper route toward destination. In Table 3 case, pedestrians were not only aware of the current location and the current direction but also read the map properly. However they had no concrete idea which route is optimal. We describe the details about the three kinds of information with the transcripts that actually observed in the experiment. In the following transcripts, ‘P:’
294
Y. Nakajima et al. Table 2. Conversation about current direction
A: So, ((should we read the map)) in this direction? ((should we go)) in this way? B: This way. A: Is this OK? B: eh? A: ((We should read this map)) in this direction, don’t you? Yes, yes, because we are walking this way now. B: This way, this way, it’s bad to read the map as it is. A: Yes, it is.
Table 3. Conversation about route selection
A: Which route? This one? B: How about downside? ((See the road heading southward)) this way? A: I agree. Because the danger zone exists over there, this way is better
means an utterance of a pedestrian and ‘N:’ means an utterance of a navigator. Number in parentheses indicates elapsed time in silence by tenth of seconds and a dot in parentheses indicates a tiny gap within or between utterances. A phrase bracketed by ‘()’ means unclear speech. A phrase bracketed by ‘(())’ means a supplement by us. The plural sentences started with ‘[’ means that they are started at the same time. ‘:’ shows sounds are stretched or drawn out (number of : indicates the length of stretching). ‘,’ means continuing intonation. ‘?’ means rising intonation and ‘.’ means closing or stopping intonation. 4.2
The Failed Case of Verbal Navigation
Questions about route selection were most frequently observed in experiment 2. A navigator tried to guide a pedestrian, but the navigator sometimes could not guide him to a proper route. Hereinafter, we discuss failure cases of verbal navigation and analyze why they failed communications. Table 4 shows an example of failed navigation caused by a navigator that could not recognize a situation of a pedestrian. In Table 4, the pedestrian asked the navigator about route selection (see line 2). In the case, the navigator could not understand the pedestrian’s situation and failed to navigate the pedestrian (line 4-11). The navigator ended the conversation with an instruction to check the current direction using the movement history. The map on a cellular phone showed an instruction to go southward but the pedestrian lost sense of direction and wrongly started going eastward as saying “I start walking randomly” (line 13). The conversation was started with the request for route guidance but the navigator could not meet the pedestrian’s demand. The navigation failed due to the difficulty for the navigator to check the pedestrian’s situation in the case.
Analysis of Pedestrian Navigation Using Cellular Phones
295
Table 4. Failure case of navigation caused by obscurity of pedestrian’s situation
1 2 3 4 5 6 7 8 9 10 11 12 13
((Pedestrian stops just after departure and ask the route to navigator)) P: Now, which direction ((to go))? (0.5) Should I turn at first, leaving the park? (1.6) N: Well:, Can you see a road (.) heading southward straight? P: ((starts walking))Southward,((look around))((turn back)) P: Well, maybe. (0.6) N: Well, hard to fi:nd? P: Yes. I can’t grasp the direction. (0.4) N: Well. (0.9) P: Oh, now (.) I start walking randomly.
A map on a cellular phone was reloaded every one minute. Location measurement and a server access took about 15 seconds on the navigation system. Consequently, the map was updated every 75 seconds. Additionally, location measurement via GPS sometimes has a gap between the current location shown on the map and the current location in the real world. Resolving the gap, the navigator was required to ask some questions to the pedestrian about the pedestrian’s situation at first. However, when the pedestrian was lost and asked the navigator for guidance, it was difficult for the navigator to get actual information from the pedestrian. The failure cases are considered to be caused by the lack of common basis between the pedestrian and the navigator. 4.3
The Successful Case of Verbal Navigation
Table 5 shows a transcript of information sharing based on a navigation map. First, the pedestrian asked the question “Does this bold black line mean this big street?” (see line 1). This question implied that the pedestrian believed to read the same map as the navigators. The navigator said “Go slightly leftward” (line 14) for guiding the pedestrian. The word “leftward” did not indicate “left side of the body (=eastward)” but “left side of the map (=westward)” in this conversation. In a word, he meant that the pedestrian had to go westward in the instruction. The pedestrian properly interpreted this confusing instruction and started heading westward without hesitation. It appears that they used the map as basis. Such a instruction was typical case when a pedestrian used maps and voice conversations. Navigations based on a movement history also occurred several times in experiment 2. As described in section 4.2, a navigation based on a movement history sometimes failed due to the difficulty of understanding the pedestrian’s situation. On the other hand, a navigation based on a navigation map often succeeded
296
Y. Nakajima et al. Table 5. Navigation based on navigation maps
1 2 3 4 5 6 7 8 9 10 11 12
P: Does this bold black line mean this big (.) street? N: Yes. (3.3) P: ((turn right)) We:[ll, N: [If you go straight, (0.4) P: Yes. N: Well, fire [disaster, you will face the fire disaster, P: [Do you mean fire disaster? P: I see. N: Go, (.) slightly leftward. P: OK.
because they can use the map as concrete common basis. Misconceptions were not likely to occur as long as a pedestrian can read maps properly. 4.4
Design Implication of Guiding Agent
We demonstrated that the kinds of information required by pedestrian were the current location, the current direction and a proper route toward a destination. Pedestrians wanted information related to the current situation but it is difficult to understand it by communication with a distant navigator. Therefore the navigation system is required to enable to check conditions of a pedestrian. Pedestrians who truly need the verbal navigation are ones who are bad at reading maps. We can support the map recognitions of such pedestrians who have a high tendency to get lost. With proper map recognition, the pedestrians can receive smooth verbal navigation from a navigator. Pedestrians could not be convinced their location even if they could get their current location via GPS. Because GPS system had little measurement error and navigation system delayed of about 75 seconds for updating a navigation map. We can improve latter point to enable mobile phone to cache the static data and to receive only dynamically changing data. Pedestrians also could not be convinced their current direction. Guiding agents calculate their current direction based on the movement histories. However, when a they just leaving a start or turning a corner, it is difficult for a guiding agent to calculate the current direction from the movement history. These days, cellular phones with electronic compasses are spreading. The system can capture precise direction of a pedestrian easily with them. In addition, pedestrians wondered which route they should select. Because there ware plural courses that can arrive at a destination. For supporting route selection, guiding agents should be equipped a function to show a route which is suited to user preferences (e.g. movement distance or a number of turns).
Analysis of Pedestrian Navigation Using Cellular Phones
5
297
Conclusion
Our approach to provide personal navigation is to build a multi-agent system that assigns one guiding agent to each human. This research attempts to demonstrate a design implication of the guiding agent. In the navigation experiments in which a pedestrian using a map on a GPS-capable cellular phone was guided by a distant navigator, we investigated the communication between them by conversation analysis method. The problems we tackled in this work are as follows. Analysis of information required by pedestrian. We examined the information required by pedestrians using the navigation system. The result indicated that pedestrians required information about the current location, the current direction and a proper route to a destination. Analysis of communication between pedestrian and navigator. Navigation maps and movement histories were used as communication basis in successful cases. When a pedestrian did not understand a map adequately, navigation sometimes failed due to lack of communication basis. In this research, we analyzed the information required by pedestrians and the limitation of remote navigation using maps and voice conversations. Future works include determining how to reflect the result of this experiment to the guiding agents and how to support people who are not good at reading maps.
Acknowledgment This work was supported by a JSPS Grant-in-Aid for Scientific Research (A) (18200009), a Grant-in-Aid for JSPS Fellows and Strategic Information and Communications R&D Promotion Programme.
References 1. Koyanagi, T., Kobayashi, Y., Miyagi, S., Yamamoto, G.: Agent server for a locationaware personalized notification service. In: Ishida, T., Gasser, L., Nakashima, H. (eds.) MMAS 2005. LNCS, vol. 3446, pp. 224–238. Springer, Heidelberg (2005) 2. Ishida, T., Nakajima, Y., Murakami, Y., Nakanishi, H.: Augmented experiment: Participatory design with multiagent simulation. In: IJCAI, pp. 1341–1346 (2007) 3. Nakajima, Y., Shiina, H., Yamane, S., Ishida, T., Yamaki, H.: Disaster evacuation guide: Using a massively multiagent server and GPS mobile phones. In: SAINT 2007, p. 2 (2007) 4. Sacks, H., Schegloff, E., Jefferson, G.: A Simplest Systematics for the Organization of Turn-Taking for Conversation. Language 50(4), 696–735 (1974) 5. Nakanishi, H., Koizumi, S., Ishida, T., Ito, H.: Transcendent communication: location-based guidance for large-scale public spaces. In: CHI 2004, pp. 655–662. ACM Press, New York (2004) 6. May, A.J., Ross, T., Bayer, S.H., Tarkiainen, M.J.: Pedestrian navigation aids: information requirements and design implications. In: PersonalUbiquitousComputer 2003, London, UK, vol. 7, pp. 331–338. Springer, Heidelberg (2003) 7. Ericsson, K.A., Simon, H.A.: Verbal reports as data. Psychological Review 87, 215– 251 (1980)
Identifying Structural Changes in Networks Generated from Agent-Based Social Simulation Models Shah Jamal Alam, Bruce Edmonds, and Ruth Meyer Centre for Policy Modelling, Manchester Metropolitan University Business School Manchester M1 3GH, United Kingdom {shah,bruce,ruth}@cfpm.org
Abstract. Finding suitable analysis techniques for networks generated from social processes is a difficult task when the population changes over time. Traditional social network analysis measures may not work in such circumstances. It is argued that agent-based social networks should not be constrained by a priori assumptions about the evolved network and/or the analysis techniques. In most agent-based social simulation models, the number of agents remains fixed throughout the simulation; this paper considers the case when this does not hold. Thus the aim of this paper is to demonstrate how the network signatures change when the agents’ population depends upon endogenous social processes. We argue for a much wider attention from the social simulation community in addressing this open research problem. Keywords: agent-based simulation, network analysis, Kolmogorov-Smirnov statistic.
Identifying Structural Changes in Networks Generated from ABSS Models
299
ABSS modellers face an even bigger challenge: characterizing the co-evolution of a dynamic network as the simulation proceeds. If agents are heterogeneous and able to join or leave the modelled system during the simulation run, the network characteristics may change completely over the course of a simulation run. Typically, snapshots of the simulated network are taken at regular time intervals. Such snapshots of the network can then be compared based on the SNA measures and other complex network characteristics. Performing a cross-sectional analysis of the network at the end of the simulation is inadequate to capture the underlying social process then. As Edmonds and Chattoe [4] have argued, in many cases an agent’s behaviour cannot just be reduced to the notion of a mere node. In the subsequent sections, we define the problem followed by a discussion on two agent-based models. We demonstrate how the use of some a priori assumptions can affect the evolved network’s structure. In the end, we suggest a potential way in uncovering the effect of dynamic population on the generated social networks.
2 Networks from Embedded Social Processes Consider a simulated network S (N, E) at the start of a simulation run, i.e. at time t=0. N is the set of nodes and E is the set of edges through which the nodes of the graph are connected. The agents are represented as nodes while the ties between them form the edges in the network. The network could be of any structure, e.g. unconnected (E = {Φ}), a small-world network [14], or a scale-free network [3]. At every time step, t: t ∈ (0,∞] new agents become a member of N, while n already present agents leave: n∈ (0, |N|].
Fig. 1. A simulation run over time, where Δ is the lag size
With each time step, the network is likely to change. Changes may happen to the size of the agent population (number of nodes in the network) and also to the formation and breaking of ties among the agents (number and position of edges). The degree distribution of the network alters depending upon the type of relations that the edges represent. In models where the joining and leaving of agents is driven by socioeconomic conditions, agents’ decision choices are constrained. The resulting network structures in such case are hard to predict in advance. The shape and size of the network at a certain time t′ could be radically different at a later time t′+Δ (where Δ is the possible time-lag of size ranging from 1, 2 … onwards; see figure 1). It is even possible that the network dissolves after some time and recovers later on.
300
S.J. Alam, B. Edmonds, and R. Meyer
Unlike physical systems, social processes tend to be modelled descriptively and validated qualitatively [8]. The evidence is often gathered through fieldwork. An individual’s relations and actions are driven by their position in society and other factors affecting the system. Social systems do not remain in a stable state and are dynamic in nature. Events changing the structure of the network may occur any time during the simulation which might be missed when using global measures. Keeping the number of nodes fixed could possibly increase the risk of type-II error or a ‘false-positive’ effect. For instance, it is possible that a network's convergence to a certain topology may be rigidly determined by the mechanism through which the agents interact. Ideally, one would apply a scheme that helps in understanding the interplay of the processes and remains robust to the changing network up to certain degree. Of course, in case the network degenerates, any such scheme is likely to fail.
3 Examples from Two Agent-Based Social Simulation Models Two different agent-based models are presented to support our proposition. The first is a model developed by Jin et al. [6] about friendship networks generated from three simple rules. The second is being developed by us based on evidence from a real case study. The two models are different with respect to their purpose and complexity and are certainly not compared in this paper. Our purpose is to highlight the likely effect of a fixed agent population and a priori assumptions about the agents’ interaction that may bias an evolved network structure. 3.1 A Model for Growing Social Network An example of simple rules resulting in the emergence of highly-clustered network structures is explained in Jin et al. [6]. The model is based on three rules: “(1) meetings take place between pairs of individuals at a rate which is high if a pair has one or more mutual friends and low otherwise; (2) acquaintances between pairs of individuals who rarely meet decay over time; (3) there is an upper limit on the number of friendships an individual can maintain.” Since the number of agents, i.e. nodes in the network remain fixed at all time, the building and breaking up of ties are solely dependent upon the algorithm used in the model. Simulating this model with 250 agents and a friendship upper limit set to 5 resulted in a quick emergence of a ‘small-world’ like structure. Once converged to a certain value, the clustering coefficient varies [13] within a small range (figure 2). We have used the implementation of the simplified model included in the Repast [11] demonstrations package. Next, we compared network snapshots P= {P0, P0+Δ, P0+2Δ…} where each element Pi represents a distinct population of agents at a particular time in the run, P0 being the population at t=0. We performed the Kolmogorov-Smirnov (K-S) test [9] for each consecutive pair (Pi, Pi+Δ) [1]. The K-S test belongs to the family of nonparametric statistics and is ‘distribution-free’. Unlike classical statistical techniques such as e.g. the t-test, there is no prior assumption about the distribution of the population from which the data is sampled. A two-population K-S-test indicates the likelihood that two datasets come from the same distribution. A comparison of the nodes degree distribution based on the
Identifying Structural Changes in Networks Generated from ABSS Models
301
1
ClusteringCoefficient
0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0
100
200
300
400
500
600
700
800
900
1000
Time steps
Fig. 2. Clustering coefficient of the friendship network [6] over a simulation run
0.9
0.9
0.6
0.6
0.3
0.3
0
0 0
100
200
300
400
500
600
700
800
900
1000
0
100
200
300
400
500
600
700
800
900
1000
Fig. 3. Y-axis: P-score of the K-S test for the degree distribution of the friendship network compared for consecutive snapshots taken at every (left) 100th tick and (right) 25th time step
K-S statistic in figure 3 shows the network stabilizing after some time. The p-score (Y-axis) that can be calculated from the KS-test can be roughly interpreted as the probability that the two data sets have the same distribution. In our context, a value 1.0 (maximum) implies that there is no change in the degree distribution of the two consecutive snapshots of the agents’ network. On the other hand, a low p-score is an indication that the network structure has changed from the previous snapshot even though, for example, the clustering coefficient remains the same. Figure 3 (left) illustrates the case when the snapshots were taken at every 100th time step. The network composition does change initially, but around the 500th time step, the degree distribution of the network stabilizes and there is no difference. Figure 3 (right) shows the comparison of networks snapshots taken at every 25th time step. As can be seen, the network starts to take shape increasingly and finally stabilizes in the distribution. Choice of a lag does influence the outcome of the statistic when comparing consecutive snapshots of the degree distribution. With a smaller lag size in figure 3 (right), one may conclude that the network’s composition stabilizes around the around 100th time step. Figure 4 shows the p-score for the K-S test statistic where we compared consecutive snapshots at the smallest possible lag size, i.e. every 5 time steps. This is due to the reason that the three basic actions ‘random meeting’, ‘neighbour meeting’ and ‘random removal’ occur at every 5th step in this model. The agents’ network grew in the initial steps and the similarity in the network configuration continued to increase thereon. In all three different granularity cases (figures 3 and 4),
S.J. Alam, B. Edmonds, and R. Meyer
P-scorefortheK olm ogorov-Sm irnovStatistic
302
0.9
0.6
0.3
0 0
5
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95
100 105 110 115 120 125
Time steps
Fig. 4. P-score of the K-S test for the degree distribution of the friendship network compared for consecutive snapshots taken at every 5th time step for the first 125 time steps of the simulation
it is clear that once the network stabilizes, its structure remains the same. One may conclude this to be a consequence of a fixed population size, where only the ties are changed. 3.2 A Model of Socioeconomic Impact of HIV/AIDS The second model is being developed based on a case study in the CAVES project1. It investigates the impact of HIV/AIDS on households and the overall community structure. The case study area is located in the Limpopo region in South Africa and is one of the most vulnerable areas lacking water, food security, jobs and other social infrastructure. Many households have female heads because the men are often living away from the house as migrant workers. State grants are the primary source of income of which a high percentage is spent on food, health and funeral costs. People try to cope with stressors via mutual help amongst neighbours, friends and extended family. A major concern is the number of orphans in the community that has increased mostly due to HIV/AIDS related adult deaths. The extended family by and large accommodates orphans of a dissolved household [15]. The model focuses on the behaviour of individual agents as well as that of households and thus attempts to take into account both the individual interactions and the decisions taken at the household level. We have adopted a multi-layer network approach to model the social networks. Individuals are represented as agents with a network of friends. Each individual is member of a household, with one of the household members acting as the household head. Households have a network of social neighbours with whom they interact. In addition, agents may join different groups like funeral clubs or savings clubs. Agents and households are created based on the available demographic data. Both endogenous and exogenous factors influence the dynamics of agent interactions. As a result, the size of the generated networks changes over time. At creation agents are assigned some random friends. With the agents joining the clubs, the size of the friendship network remains dynamic. A high prevalence of HIV/AIDS affects the health of those who get infected resulting in a high increase death. Each individual 1
Identifying Structural Changes in Networks Generated from ABSS Models
303
agent is a member of a household. Households in turn form clusters which represent the extended family. This comes into play when a household dissolves when the parents have died; an accommodating household is looked for. This search uses the family hierarchy to determine the nearest living relative who is able to accommodate the surviving dependants [2]. The simulation runs evaluated in this paper were initialized with 30 households and about 150 adult and senior agents. These numbers were chosen because the online tool used in calculating the Kolmogorov-Smirnov p-score allowed only a limited data set size [7]. For future work, we will apply standard statistical software for a larger number of agents. In this setup, only the friendship network of adult and senior agents has been considered. We added the three rules for growing the friendship network from the model by Jin et al. [6]. In constrast to the latter model, the size of the population is not fixed – agents die and are born over the course of the simulation. Agents’ expected life was sampled from a normal distribution with a mean of 75 years and standard deviation 10 years. However, the actual lifespan of an agent may be shorter due to declining health. This is influenced both by the socioeconomic status of a household, which affects the food intake of its members, and the prevalence of HIV/AIDS. The birth rate was set to 15 children per 1000 individuals of a population. The size and the network composition resulting from this model cannot be assumed in advance. To investigate the impact of HIV/AIDS, or in other words the impact of a rapidly decreasing agent population on the evolution of the friendship network we performed two batches of simulation runs, one without prevalence of HIV/AIDS and one with a prevalence of 14%. Agents were randomly assigned 1 to 5 friends thus using the maximum limit as in the model described in the previous section. The first batch comprised of three simulation runs with slightly different settings regarding the rules governing the formation and breaking of friendship links. Figure 5 (left) shows the clustering coefficient of the agents’ friendship network for the three different cases. In the first case (Case-I), friendships are formed by two of the three rules from the Jin et al. model [6], the neighbour meeting and the random meeting rule. Agents also make friends when they join the same club. Friendship ties only break when agents die. Case-II incorporated the random removal rule in addition, whereas for Case-III, only the three rules from [6] were used and the same club membership was ignored In all the three cases, the network remained sparse with a small clustering coefficient. Adding the random removal rule did influence the abundance of links initially, but was not significant in the long run. The relative changes in the friendship network’s degree distribution were more evident when the Kolmogorov-Smirnov test was applied to all the three cases, as in the previous model. While a global property of a network, here the clustering coefficient, showed a similar pattern, networks actually changed in different ways for each case as shown in figure 5 (right). In the next step, the simulation as in Case-III was run, but this time with an HIV/AIDS prevalence of 14%. The rapidly declining agent population causes the clustering coefficient (in figure 7; left) to decline much faster as compared to the three previous cases in figure 5. Most agents die around the 600th time step.
304
S.J. Alam, B. Edmonds, and R. Meyer
The p-score of the K-S statistic drops between time periods 350-450 indicating significant change occurring in the network. This happens due to increasing deaths in the system, which affects the degree distribution of the agents’ friendship network. As the network becomes sparse after the 500th step, the corresponding p-value shows no change at all, followed by a sharp decline when the entire community structure collapses. 1.2
P-score of Kolmogorov-Smirnov Statistic
0.45 0.4
Clustering Coefficient
0.35 0.3 0.25 0.2 0.15 0.1 0.05
1
0.8
0.6
0.4
0.2
0
0 0
50
100
150
200
250
300
350
400
450
500
550 600 650 700
750
0
800
50
100 150 200
250 300 350
Case-I
400
450 500 550
600 650 700
750 800
Time steps
Time steps
Case-II
Case-III
Case-I
Case-II
Case-III
Fig. 5. (left) Clustering coefficient of the agents’ friendship network; (right) P-score of the K-S test for the degree distribution taken at every 50th time step for the three different simulation settings
To further investigate whether the changes in the number of agents does really make an impact on the network signatures, we ran simulations for three additional cases without prevalence of HIV/AIDS. In Case-I, no birth of an agent was possible, so the agent population slowly declined Case-II used the opposite scenario with only births but no deaths happening during the simulation run. For Case-III we kept the initial population of agents constant throughout the run. 1.2
P-score of Kolmogorov-Smirnov Statistic
0.35
Clustering Coefficient
0.3
0.25
0.2
0.15
0.1
0.05
0.9
0.6
0.3
0
0 0
50
100
150
200
250
300
350
400
450
Time step
500
550
600
650
700
750
800
50
100
150 200
250
300 350
400 450
500 550
600
650 700
750 800
Time steps
Fig. 6. (left) Clustering coefficient of the agents’ friendship network; (right) P-score of the K-S test for the degree distribution taken consecutively at every 50th time step, with HIV/AIDS prevalence
Figure 7 (left) shows the clustering coefficients for these three cases. For the first case we get a similar pattern as in figure 6 (left). Whereas in Case-I no birth was allowed, in the latter case agents’ fertility had declined due to the HIV/AIDS
Identifying Structural Changes in Networks Generated from ABSS Models
305
prevalence, thus equally constraining births. Case-II follows a similar pattern as the simulations displayed in figure 5, where the population is equally growing due to the high birth rate. Case-III however, shows a high clustering coefficient, well-above 0.5 after some initial steps. This is very similar to the clustering coefficient obtained for the Jin et al. model in figure 2. The reason for achieving a so-called ‘small-world’ effect is that the agent population remained the same in this case. The p-score for the Kolmogorov-Smirnov (K-S) statistic for the respective three cases is shown in figure 8 (right). Similar network configurations may be observed for Case-I and the HIV/AIDS prevalence scenario shown in figure 7 (right). The K-S statistic for Case-II keeps changing over time which is due to the birth of agents during the simulation. For Case-III, the signature is again almost identical to the Jin et al. model run as shown in figure 3. Once the agents’ friendship network stabilizes it remains the same for the rest of the simulation run. 0.6
P-score of Kolmogorov-Smirnov Statistic
1.2
Clustering Coefficient
0.5
0.4
0.3
0.2
0.1
0
0.9
0.6
0.3
0 0
50
100
150 200
250 300
350
400 450
500 550
600 650
700 750
800
50
100
150 200
250
300 350
Time steps Case-I
Case-II
400 450
500 550
600
650 700
750 800
Time steps
Case-III
Case-I
Case-II
Case-III
Fig. 7. (left) Clustering coefficient of the agents’ friendship network; (right) P-score of the K-S test for the degree distribution taken consecutively at every 50th time step for the three different simulation settings
4 Discussion and Outlook A very few papers about social network models assume that the number of agents will change over time. Models of network evolution, e.g. the Barabasi-Albert model of preferential attachments provide the mechanism whereby the number of agents monotonically increases (but does not decrease). Even so, such network evolution is deterministic and can be predicted in advance. On the other hand, agent-based models where an agent’s membership in the network is constrained by its local interaction processes, the population size cannot be guessed for larger time lags in advance. In this paper, we have discussed the issue of changes in the network structural signatures in case of a dynamical population. We have examined the stability of network measures on two different agent-based simulation models where networks evolve over time. The network measures, i.e. the degree distribution and the clustering coefficient are the most prevalent measures in the network analysis and we have applied them to the simulated networks as demonstration. It is shown that even the robust measures behave differently under varying population during the simulation despite similar generative mechanisms. Nevertheless, the instability of networks measures
306
S.J. Alam, B. Edmonds, and R. Meyer
can be generalized to more advanced network measures that are used in analyzing the simulated networks. The Kolmogorov-Smirnov (K-S) test is used to test if there is a difference in the distribution of two populations. The K-S test provides a means to look into the stability of measures on pure statistical basis and thus can be applied to any agent-based simulation case study. Previously, Kossinets and Watts [16] have used the K-S statistic for a longitudinal empirical friendship network of e-mail logs for fixed sized time windows. We have, in this paper, considered different network snapshots of the same simulation run as separate populations and compared the measures them using the KS test. Our main motivation has been to look for techniques that can help keeping track of the network structure over the course of a simulation run. Observing the pscore for consecutive snapshots can indicate whether the network structure has changed or not. A high p-score is an indicator that there is no evidence from the distributions that the network has changed – a difference is possible but this is unlikely. Moreover, significant differences in the p-score could be one possible indicator in selecting a set of snapshots for further analysis. It is interesting to note that the network measures such as the clustering coefficients vary as the denominator changes. The measure remains stable in the dynamical networks, when the isolated or the unconnected agents (or nodes) are not considered. The removal (or death) of an agent in the population may alter the number of unconnected nodes in the network. As a result, the transitivity in the network no longer remains the same. The point we want to make here is that for real case study problems, it might be a good idea to check whether the evolved network is due to certain procedures that specifically lead to a particular topology. Networks resulting from social processes may or may not produce topologies such as the scale-free or a smallworld topology. We are working towards further results to support that this point holds for all the agent-based simulations in general and can therefore call for much wider attention of the MAS community. Acknowledgments. Thanks to our colleagues at the Centre for Policy Modelling for their useful comments. This work is supported under the EU FP6 Project CAVES.
References 1. Alam, S.J., Meyer, R., Norling, E.: Agent-based Model of Impact of Socioeconomic Stressors: A Dynamic Network Perspective. In: Sixth Intl. Joint Conf. on Autonomous Agents and Multiagent Systems, Hawaii, USA, pp. 1079–1081 (2007) 2. Alam, S.J., Meyer, R., Norling, E.: Using Agent-Based Modeling to Understand the Impact of HIV/AIDS in the Context of Socio-Economic Stressors. CPM-06-167 (2006), http://cfpm.org/cpmrep167.html 3. Albert, R., Barabasi, A.: Statistical Mechanics of Complex Networks. Rev. Mod. Phys. 74 (2002) 4. Edmonds, B., Chattoe, E.: When Simple Measures Fail: Characterising Social Networks Using Simulation. Social Network Analysis: Advances and Empirical Applications Forum. Oxford, UK (2005), http://cfpm.org/cpmrep158.html 5. Faust, K.: Comparing Social Networks: Size, Density, and Local Structure. Metodološki zveski 3(2), 185–216 (2006)
Identifying Structural Changes in Networks Generated from ABSS Models
307
6. Jin, E.M., Girvan, M., Newman, M.E.J.: Structure of growing social networks. Phys. Rev. E 64, 4 (2001) 7. Kirkman, T.W.: Statistics to Use (1998), http://www.physics.csbsju.edu/stats/ 8. Moss, S., Edmonds, B.: Sociology and Simulation: Statistical and Qualitative CrossValidation. AJS 110(4), 1095–1131 (2005) 9. Neave, H., McConwa, K.: Distribution free methods. OUP (1987) 10. Newman, M.E.J.: The structure and function of complex networks. SIAM Review 45, 167–256 (2004) 11. North, M.J., Collier, N.T., Vos, J.R.: Experiences Creating Three Implementations of the Repast Agent Modeling Toolkit. ACM Trans. on Modeling and Computer Simulations 16(1), 1–25 (2006) 12. Pujol, J.M., et al.: How Can Social Networks Ever Become Complex? Modelling the Emergence of Complex Networks from Local Social Exchanges. JASSS 8(4), 12 (2005), http://jasss.soc.surrey.ac.uk/8/4/12.html 13. Wasserman, S., Faust, K.: Social Network Analysis: Methods and Applications. Cambridge University Press, Cambridge (1994) 14. Watts, D.J., Strogatz, S.: Collective Dynamics of ‘small world’ networks. Nature 393, 440–442 (1998) 15. Ziervogel, G., et al.: Adapting to climate, water and health stresses: insights from Sekhukhune, South Africa. Stokholm Environment Institute (Oxford) (2006) 16. Kossinets, G., Watts, D.J.: Empirical analysis of an evolving social network. Science 311, 88–90 (2006)
Appendix: A Scheme for Comparing Changes in the Network Structure During a Simulation Run I. Choose a lag of size Δ for the comparison of network snapshots at different time steps. II. For all time lags ti (iÅ 0,Δ, 2Δ,…), obtain a corresponding series of the node degrees, Pi. (We call this the profile) III. For all consecutive pairs (Pi, Pi+Δ) (iÅ 0,Δ, 2Δ,…), compute the p-score for the Kolmogorov-Smirnov (K-S) test for 2-populations. The K-S test gives an indication of the likelihood that the two series are from the distribution. A low p-score indicates that it is unlikely that the two series are from the same distribution, and hence that the degree distributions and hence the network type differs in some respect. A high p-score would mean that there is no evidence that the distributions are different and hence is a fairly tough test of network similarity. However it is still possible that the networks do differ, albeit in subtle ways. The webpage used for this test is: http://www.physics.csbsju.edu/stats/
Multi-agent Simulation of Linguistic Processes: A NEPs Perspective Gemma Bel-Enguix and M. Dolores Jim´enez-L´opez Research Group on Mathematical Linguistics Rovira i Virgili University Pl. Imperial Tarraco, 1, 43005 Tarragona, Spain {gemma.bel,mariadolores.jimenez}@urv.cat
Abstract. Networks of Evolutionary Processors (NEPs) –introduced in Castellanos et al. (2001)– are a new computing mechanism directly inspired in the behaviour of cell populations. NEPs, as well as generating devices, can be considered as bio-inspired context-sensitive multi-agent systems. This paper analyzes the agential features of NEPs and their special responses to the environment, which make this system specially suitable for description of Natural Language.
1
Introduction
In the last decades, biology has become a rich source of models for other sciences. The knowledge of the behaviour of nature has influenced a number of areas such as artificial intelligence, mathematics or theoretical computer science, giving rise to new perspectives in research, like natural computing. Neural networks [17], genetic algorithms [29], ants’ theory [11] and L systems [27] are just some examples of this eclosion of new bio-inspired computational paradigms. One of the most developed lines of research in natural computing is the named molecular computing, a model based on molecular biology, which arose mainly after Adleman’s work [1]. An active area in molecular computing is DNA computing [26], inspired in the way DNA performs operations to generate, replicate or change the configuration of the strings. Furthermore, during the last years, systems biology and cellular biology have achieved an important development. These advances have provided new models for computer science. One of them is cellular computing, that emphasizes the concept of microbiological populations as well as the equilibrium of the devices and the relationships between the elements. P systems [25] can be considered an example of this emerging paradigm. On the other hand, natural computing has evolved from the first numeric models –like neural networks– to symbolic models –like cellular computing– which are closer to multi-agent systems. Networks of Evolutionary Processors (NEPs) [3] are inspired in both, in bio cellular models and in basic structures for parallel and distributed symbolic processing. A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 308–318, 2009. c Springer-Verlag Berlin Heidelberg 2009
Multi-agent Simulation of Linguistic Processes: A NEPs Perspective
309
The cellular basis of NEPs relate them with P systems, especially with Tissue P systems [24], a theory in the area of membrane computing whose biological referent is the structure and behaviour of multicellular organisms. In tissue P systems, cells form a multitude of different associations performing various functions. From the computational point of view, NEPs are related to the Connection Machine [18] and the Logic Flow paradigm [12]. Another important theoretical relationship of NEPs is the theory of Grammar and Eco-grammar Systems [7], which provide a syntactical framework for communities of evolving agents and their interrelated environment. With all this background and theoretical connections, it is easy to understand how NEPs can be described as bio-inspired context-sensitive multi-agent systems. Many disciplines are needed of these types of models that are able to support a biological framework in a collaborative environment. The conjunction of these features allows applying the system to a number of areas, beyond generation and recognition in formal language theory. Linguistics is one of the fields with a lack of biological models and with a clear suitability for agential approaches. We introduce in this paper a method that aims not only to be implemented, but also to be able to generate a computational simulation of a number of language processes. Therefore, what we are not suggesting a new multi-agent theory that can help in the resolution of several open questions in linguistics, but a new method of linguistic research and simulation, which is based in a bio-cellular model with multi-agent capabilities. The introduction of such model can help to change the perspective of theoretical linguistics, causing that implementation could become a new central tool for the development of linguistic formal theories. In the sequel, we introduce the computational definition of NEPs (Section 2), highlighting some of the most important features of this type of multi-agent systems (Section 3). After explaining what is the signification of such traits for linguistics (Section 4), we show a simple example of an implementation of the morphological component of language (Section 5).
2
NEPs: Definition
As it has been already pointed out, Networks of Evolutionary Processors are a new computing mechanism directly inspired in the behaviour of cell populations. NEPs can be defined as systems consisting of several devices whose communication is regulated by an underlying graph. Such devices, which are an abstract formalization of cells, are described by a set of words (DNA) evolving by mutations, according to some predefined rules. Their outcome travels to the other nodes if they accept it after passing a filtering process. At the end of the process, only the cells with correct strings will survive. This mechanism allows the specialization of each processor. First precedents of NEPs as generating devices can be found in [10] and [9]. The topic was introduced in [3] and [22], and further developed in [2], [4], [8], and [23]. A new approach to networks of evolutionary processors as accepting devices has started in [20]. A scheme of the basic architecture of NEPs is shown in Figure 1.
310
G. Bel-Enguix and M.D. Jim´enez-L´ opez
N3
N4
N5
N2
N1
N0
Fig. 1. Graph Structure
Following the basic definition of NEPs, a Network of Evolutionary Processors of size n is a construct: Γ = (V, N1 , N2 , ..., Nn , G), where: – V is an alphabet and for each 1 ≤ i ≤ n, – Ni = (Mi , Ai , P Ii , P Oi ) is the i-th evolutionary node processor of the network. The parameters of every processor are: • Mi is a finite set of evolution rules of one of the following forms only i. a → b, a, b ∈ V (substitution rules), ii. a → e, a ∈ V (deletion rules), iii. e → a, a ∈ V (insertion rules). • Ai is a finite set of strings over V . The set Ai is the set of initial strings in the i-th node. • P Ii and P Oi are subsets of V ∗ representing the input and the output filter, respectively. These filters are defined by the membership condition, namely a string w ∈ V ∗ can pass the input filter (the output filter) if w ∈ P Ii (w ∈ P Oi ). – G = ({N1 , N2 , . . . , Nn }, E) is an undirected graph called the underlying graph of the network. The edges of G, that is the elements of E, are given in the form of sets of two nodes. The complete graph with n vertices is denoted by Kn . Configuration of a NEP is an n-tuple C = (L1 , L2 , . . . , Ln ), with Li ⊆ V ∗ for all 1 ≤ i ≤ n. It represents the sets of strings which are present in any node at a given moment. A given configuration of a NEP can change either by an evolutionary step or by a communicating step. When changing by an evolutionary step, each component Li of the configuration is changed in accordance with the evolutionary rules associated with the node i. The change in the configuration by an evolutionary step is written as C1 ⇒ C2 . When changing by a communication step, each node processor Ni sends all copies of the strings it has which are able to pass its output filter to all the node processors connected to Ni and receives all copies of the strings sent by any node processor connected with Ni , if they can pass its input filter. The change in the configuration by a communication step is written as C1 C2 .
Multi-agent Simulation of Linguistic Processes: A NEPs Perspective
311
This formal definition of NEP, which has been developed in the last years by theoretical computer science, can provide a good framework for attempting a new description and formalization of linguistics. In the sequel, we discuss several features of NEPs, explaining why they are suitable for linguistics.
3
NEPs: Features
From the definition above, and taking into account their intrinsic characteristics, NEPs can be defined as modular evolutionary environmental multi-agent systems. They take the multiagent and modular features from classical theoretical computer science, while the environmental and evolutionary characteristics are typical of bio-inspired models. We describe the chief traits of such multi-agent systems as follows: – Modular: NEPs are distributed systems of contributing nodes, each one of them carrying out just one type of operation. Every node should be defined depending on the specific domain we aim to tackle. – Multi-level modularity: Every specialized processor can be modular as well. This architecture is quite suitable for linguistics, a science whose units in every level are obtained by gathering smaller segments up to reach, in the highest one, the most complex constructs. Such path, in linguistics, is reversible. In the present paper, we show an example of the morphological module understanding its internal structure as another NEP. – Evolutionary: agents/nodes can change their definition during the computation. Such modifications can be caused by the context, by the elements inside or by the learning of the modules. – Environmental: By this feature we explain that NEPs are systems-in-context. Context has an important influence in the behaviour of nodes/agents. By context of one node we understand: a) the other processors in the same system, and b) the environment in which the system is placed. Both elements are crucial for describing the state of an element in a given moment and its evolution in a time sequence. Inside of the construct, every agent is autonomous, specialized, contextinteractive and learning-capable. These features can be described as follows: – Autonomous: Autonomy and specialization are strongly related in NEPs. Even if every agent depends on the other to accomplish the final objective, in order to improve the final result they can change the strategy and the own sub-goals. – Specialized: Every processor is designed for a specific task, in a way that the final success of the system depends on the correct working of every one of the agents. – Context-Interactive: There is a constant contextualization and redefinition of agent capabilities during the computation. In our model, we postulate that language and context modify each other, in a way that utterances imply an
312
G. Bel-Enguix and M.D. Jim´enez-L´ opez
action that modifies the ecosystem and has a global incidence. This is the way how NEPs can model evolution in both, the agents and the environment. The interaction with context implies three different abilities of agents. They are: • Context-sensitive: Having a broad knowledge of the configuration of the environment and the other agents of the system • Context-reactive: Changing their goal and behaviour depending on the configuration of the environment. • Context-modifier: The productions and behaviour of a given cell are able to modify the general context. – Learning-Capable: Evolution in nodes can be caused by learning. Unsuccessful computations can change the configuration or the interaction between processors. Between the processors, there are social competences, specified by a GraphSupported structure and Filtering-Regulated Communication and Coordination. – Communication. Agents can interact in the consecution of the same goal, work for their own interests or even isolate. – Coordination. It is necessary to define protocols for synchronizing the communication. The two mechanisms that regulate the interaction in the multi-agent system –the graph structure and the filters– have an important role in both the specialization of the processors and the correct functioning of the device. These features are explained as follows: – Filter-regulated: By means of input and output filters. • The objective of the input filter is to control the information or structures entering the node. This helps to the specialization, by the selection of the strings/structures the node can process, and protects the module from possible harmful items. • By the output filter the node selects the information it wants to share and also when it wants to share it. – Graph-supported: The edges of the graph govern the communication. Non connected graphs do not communicate. This mechanism decreases the complexity of the working of the whole system and allows interaction between processors. However, communication between nodes and the environment is not graphsupported. We establish that the agents know, in every moment, what is the configuration of the general cytosolic area where the system is placed. Finally, in what refers to the functioning of the NEPs, two main features deserve to be highlighted. – Emergence: The behaviour of the nodes is not necessarily pre-established. It depends on both, the evolution of the system and the conditions of the environment.
Multi-agent Simulation of Linguistic Processes: A NEPs Perspective
313
– Parallelism: Different tasks can be performed at the same time by different processors. Some of linguistic processes, as well as language generation in general are considered to be parallel. For apparently sequential interactions (i.e. dialogue) parallelism allows working with multi-modal exchanges.
4
NEPs for Modelling Linguistics
Because of the main features we have mentioned above, NEPs seems to be a quite suitable model for tackling linguistics. Human language is traditionally a difficult entity to deal with, to approach by formal methods, and to describe. One of the main problems of verbal language is that it is generated in the brain, and there is an important lack of knowledge of the mental processes the mind undergoes to, finally, bring about a sentence. While expecting new and important advances in neuro-science and neurolanguage, we are forced to use the models that seem to fit better to language generation and recognition. Such models are usually related to cognitive theories, that have also inspired different perspectives in computer science. Up to now, one of the hypotheses that have produced the best approximations to natural language processing, as well as most of cognitive processes, is modularity. Even though the idea of modularity has been implicit in cognitive science for a long time, it is with the publication of The Modularity of Mind (Fodor, [15]) when those implicit ideas that had been current over the previous two decades crystallized into a single recognizable hypothesis. In general, modular theories in cognitive science propose a number of independent but interacting cognitive ‘modules’ that are responsible for each cognitive domain. Specific arrangement of those modules usually varies in each theory, but generally each mental module encapsulates a definable higher mental function. Fodor’s theory is not by far the only one about modularity of mind. In fact, in the 1980s there started a new trend represented by authors such as Chomsky [6] and Jackendoff [19]. The theory of modularity is also present in linguistic approaches. In fact, the idea of having a system made up of several independent components (syntax, semantics, phonology, morphology, etc.) has been often assumed in linguistics. We may cite several modular approaches in Linguistics, from Chomsky’s Generative Grammar [5] and Halle’s Morphology [16] to Autolexical Syntax [28] or Jackendoff’s view of the Architecture of Language Faculty [19] just to name a few. One of the main advantages of modular grammar is that they can reduce delays imposed by monolithic or non-modular grammars. This reduction of delays is due to the fact that, in modular grammars, subsystems of grammatical principles can be applied independently of each other and, sometimes, in a parallel fashion. Therefore, taking the four main modules usually considered in linguistics – syntax, semantics, phonology, morphology– and considering the edges of the underlying graph as way for communicating or not communicating, a simple scheme of a “language generation NEP” can be drawn where N0 stands for phonetics, N1 represents morphology, N2 is syntax and N3 refers to semantics.
314
G. Bel-Enguix and M.D. Jim´enez-L´ opez
The semantic node is only communicated to the phonological and the syntactic ones because it seems there is not interaction between semantics and phonetics. Modularity appears in linguistics together with a multi-level specialized structure. This implies that each one of the main modules could have also a modular structure [30,13,14]. A good example of that is the Chomsky’s theory of Government and Binding [5], where several components can be found in syntax: government theory module, binding theory module, bounding theory, move-alpha, etc. In the same theory, there are several syntactic levels (D-structure, S-structure, and LF). Eventhough the first formalization of NEPs did not expect those type of constructions, there is no theoretical reason that prevents them. Multi-level modular NEPs could be designed in order to simulate the whole process of sentence generation, for example. If NEPs are modular multi-agent systems, they have to communicate in order to exchange results. If we defend a specialized approach, we have to establish ways of interaction between the nodes of the computing device. Since the underlying graph of the systems is not necessarily complete, communication between two nodes can be regulated by the existence –or not– of an edge that unites them. Several protocols of creation or deletion of nodes have to be formalized as well, in order to increase the capacity of evolution and adaptation of the network. Another chief problem for the formalization of human verbal language is its changing nature. Not only words change through the time. Also rules, meaning and even phonemes –the first components of language– can take different shapes during the process of computation. Formal models based only in mathematical (an extremely stable) language present a lack of flexibility to describe natural language. Biological models seem to be better to this task, since biological entities share with languages the concept of “evolution” as one of the main features of the system. A number of authors (McMahon [21] have adopted – or adapted – biological evolution itself in order to construct an evolutionary theory of language change. From this perspective, NEPs offer enough flexibility to model any change at any moment in any part of the system. Besides, as a bio-inspired method of computation, they have the capability of simulating natural evolution in a highly pertinent and specialized way. However, many unexpected changes or strange uses of rules in language, are not caused by evolution, but by social constraints: historical context, interaction with other users, secret goals of the speakers... These features can be modelled by what we call context-interactivity. It is in this aspect where the classical theory of formal languages has achieved the worst results for dealing with natural language. Some linguistic disciplines, as pragmatics or semantics, are context-driven areas, where the same utterance has different meanings in different contexts. To model such variation, a system with a good definition of environment is needed. NEPs, which are based in colonies of cell populations that behave according to the configuration of the environment, seem to be able to offer some kind of solution to approach formal semantics and formal pragmatics from a natural computing perspective.
Multi-agent Simulation of Linguistic Processes: A NEPs Perspective
315
Moreover, linguistics has always been studied from a linear and sequential point of view. Indeed, the sound is produced in a sequential way, but many studies could suggest that the production could not be sequential. This is, for example, what the Chomsky’s idea of rewriting seems to suggest. Furthermore, some linguistic theories, like autolexical syntax (Sadock, [28]), are defined by their authors as parallel, indicating that “natural-language expressions are organized along a number of simultaneous informational dimensions”. Finally, the multimodal approach to communication, where not just production, but also gestures, vision and supra-segmental features of sounds have to be tackled, refers to a parallel way of processing. NEPs allow to work in the same goal, but in different tasks, to each one of the modules. The autonomy of every processor and the possible miscoordination between them can also give account of several problems of speech.
5
An Example: Halle’s Lexicalist Hypothesis
We have shown how the features that have been highlighted in the design of NEPs are also conditions that should provide a robust framework for linguistics. Many linguistic theories are based in these principles and, therefore, should be able of being formalized under these parameters. So, to illustrate our proposal, we introduce an example of application of NEPs to language processing in a modular multi-level way. To do so, we take a classical and very fruitful proposal of Morris Halle about morphology. Halle [16] postulated the lexicalist theory for word formation. A rough explanation of the theory could say that words are obtained combining morphemes by means of word formation rules (WFR). But not every word generated by this method is good. There are actual words (un-happy-ness), potential words (un-dog-ly) and gaps (un-good-ly). To distinguish between them, it is necessary an output filter for these productions. But once filtered, these words could have to be modified by syntactic or phonetic rules (un-happi-ness). At the end of the process, we expect to get only the correct words of the language. Keeping in mind these elements, we can define a second level modular system for Halle’s lexicalist hypothesis as follows: Γ = (V, N1 , N2 , N3 , G, E), where: – – – – –
V is the set of morphemes of the language. N1 = (WFR, ∅, V ∗ , W ∈ V ∗ ). N2 = (SR, ∅, W, W s ∈ V ∗ ∪ W p ∈ V ∗ ). N3 = (PR, ∅, W s ∈ V ∗ , W p ∈ V ∗ ). G = (V, E) is the underlying graph, being: • V = {N1 , N2 , N3 }. • E = {N1 N2 , N2 N3 , N3 N1 }. – E ⊂ V is a set of morphemes over V .
This formalization has added something to the basic computational definition of NEP, the environment (E), in order to increase the power of the context. In our description WFR stands for Word Formation Rules, SR for Syntactic Rules and PR for Phonetic Rules. It is clear that the first node will compose words, but not all of them can be considered words of the language. A first filtering (W) does not allow ‘potential words’ and ‘gaps’ to leave the processor N1 . Some of the strings that are able to pass the filter will need a new processing in the syntactic module N2 and others even a third processing in the phonetic module. The configuration of the context establishes a set of morphemes that will be the initial input of the system. Every node in the system is, by definition, an initial node. The input (morphemes) is taken from the environment, and every processor could do that. However, in this NEP, only the filter of N1 allows the morphemes to enter the processor. Both, N2 and N3 are final nodes, and their productions, if correct, can leave the morphologic module and go directly to the syntactic module. Halle’s Hypothesis has been discussed and improved during thirty years, but it is a very good example of modularity, interaction between modules, and filtering. The NEP’s design is shown in Figure 2. The example is simple, but it can illustrate the process and suggest more complex architectures for the future.
6
Final Remarks
The paper presents a quite new bio-inspired multi-agent method of computation, NEPs. After explaining some of the main features of the system, we remark its suitability for tackling several fields of linguistics that have been difficult to formalize. The idea of a multi-agent modular approach to syntax has several exponents in the last years. But it is quite unusual to aim a formal and modular approach to semantics and pragmatics. We think NEPs provide the theoretical background to do that.
Multi-agent Simulation of Linguistic Processes: A NEPs Perspective
317
The coincidences between several structures of language and biology allow us to take advantage of the bio-inspired models formalized by theoretical computer science. Moreover, the multi-agent capabilities of these models make them a suitable tool for simulating the processes of generation and recognition in natural language. From a theoretical point of view, maybe NEPs cannot be used to explain how language is generated (this is part of neurology and biology), but they can help to implement and simulate such processes. The results achieved can be a way to falsate, test and debug the formal weakness of current hypothesis as well as provide feed-back to suggest new lines of research. So, the paper has two main interests: the first one involves language and computability; the second one is related to language and biology. First, the theory we present can show several methods for artificial intelligence to achieve an implementable formalization of natural language, capturing several aspects (i.e. context, evolution, meaning, etc.) that traditionally have been hard to manage for natural language processing. On the other hand, it can be a way to improve, by means of simulation of mental processes, the research on how language is generated and recognized. The main idea for the future is to start an implementation of NEPs, in order to prove that the computational simulation of language by this mechanism is possible. For our purposes, we suggest to perform the following improvements in the basic design of linguistic NEPs: a) to strengthen the role of environment in the system; b) to take advantage of the idea of modularity, which is related to the architecture of NEPs and to symbolic conceptions of cognitive processes; c) to develop communication protocols using the underlying graph.
References 1. Adleman, L.M.: Molecular Computation of Solutions to Combinatorial problems. Science 226, 1021–1024 (1994) 2. Castellanos, J., Mart´ın-Vide, C., Mitrana, V., Sempere, J.M.: Solving NP-complet Problems with Networks of Evolutionary Processors. In: Mira, J., Prieto, A.G. (eds.) IWANN 2001. LNCS, vol. 2084, pp. 621–628. Springer, Heidelberg (2001) 3. Castellanos, J., Mart´ın-Vide, C., Mitrana, V., Sempere, J.M.: Networks of Evolutionary Processors. Acta Informatica 39, 517–529 (2003) 4. Castellanos, J., Leupold, P., Mitrana, V.: On the Size Complexity of Hybrid Networks of Evolutionary Processors. Theoretical Computer Science 330(2), 205–220 (2005) 5. Chomsky, N.: Lectures on Government and Binding. The Pisa Lectures. Foris Publications, Dordrech (1981) 6. Chomsky, N.: Modular Approaches to the Study of Mind. San Diego State University Press (1984) 7. Csuhaj-Varj´ u, E., Dassow, J., Kelemen, J., P˘ aun, G.: Grammar Systems. Gordon and Breach, London (1993) 8. Csuhaj-Varj´ u, E., Mart´ın-Vide, C., Mitrana, V.: Hybrid Networks of Evolutionary Processors are Computationally Complete. Acta Informatica 41/4, 257–272 (2005)
318
G. Bel-Enguix and M.D. Jim´enez-L´ opez
9. Csuhaj-Varj´ u, E., Mitrana, V.: Evolutionary Systems: A Language Generating Device Inspired by Evolving Communities of Cells. Acta Informatica 36, 913–926 (2000) 10. Csuhaj-Varj´ u, E., Salomaa, A.: Networks of Parallel Language Processors. In: P˘ aun, G., Salomaa, A. (eds.) New Trends in Formal Languages. LNCS, vol. 1218, pp. 299–318. Springer, Heidelberg (1997) 11. Dorigo, M., St¨ utzle, T.: Ant Colony Optimization. MIT Press, Cambridge (2004) 12. Errico, L., Jesshope, C.: Towards a New Architecture for Symbolic Processing. In: Plander, I. (ed.) Artificial Intelligence and Information-Control Systems of Robots 1994, pp. 31–40. World Sci. Publ., Singapore (1994) 13. Everaert, M., Evers, A., Hybreqts, R., Trommelent, M. (eds.): Morphology and Modularity. In: Honour of Henk Schultink, Publications in Language Sciences 29, Foris (1988) 14. Farmer, A.K.: Modularity in Syntax: A Study of Japanese and English. MIT Press, Cambridge (1984) 15. Fodor, J.: The Modularity of Mind. MIT Press, Cambridge (1983) 16. Halle, M.: Prolegomena to a Theory of Word Formation. Linguistic Inquiry 4, 3–16 (1973) 17. Haykin, S.: Neural Networks: A Comprehensive Foundation. Prentice Hall, Englewood Cliffs (1999) 18. Hillis, W.D.: The Connection Machine. MIT Press, Cambridge (1985) 19. Jackendoff, R.: The Architecture of Language Faculty. MIT Press, Cambridge (1997) 20. Margenstern, M., Mitrana, V., P´erez-Jim´enez, M.: Accepting Hybrid Networks of Evolutionary Processors. In: Ferreti, C., Mauri, G., Zandron, C. (eds.) DNA 10. Preliminary Proceedings, Milan, University of Milano-Biccoca, pp. 107–117 (2004) 21. McMahon, A.: Understanding Language Change. Cambridge University Press, Cambridge (1994) 22. Mart´ın-Vide, C., Mitrana, V., P´erez-Jim´enez, M., Sancho-Caparrini, F.: Hybrid Networks of Evolutionary Processors. In: Cant´ u-Paz, E., et al. (eds.) GECCO 2003. LNCS, vol. 2723, pp. 401–412. Springer, Heidelberg (2003) 23. Mart´ın-Vide, C., Mitrana, V.: Networks of Evolutionary Processors: Results and Perspectives. In: Molecular Computational Models: Unconventional Approaches. Idea Group Publishing, Hershey (2004) 24. Mart´ın-Vide, C., Pazos, J., P˘ aun, G., Rodr´ıguez-Pat´ on, A.: A new class of symbolic abstract neural nets: Tissue P systems. In: H. Ibarra, O., Zhang, L. (eds.) COCOON 2002. LNCS, vol. 2387, pp. 290–299. Springer, Heidelberg (2002) 25. Pˇ aun, G.: Computing with Membranes. Journal of Computer and System Sciences 61, 108–143 (2000) 26. Pˇ aun, G., Rozenberg, G., Salomaa, A.: DNA Computing. New Computing Paradigms. Springer, Berlin (1998) 27. Rozenberg, G., Salomaa, A.: The mathematical theory of L Systems. Academic Press, New York (1980) 28. Sadock, J.M.: Autolexical Syntax. A Theory of Parallel Grammatical Representations. University of Chicago Press (1991) 29. Schmitt, L.: Theory of Genetic Algorithms. Theoretical Computer Science 259, 1–61 (2001) 30. Weinberg, A.: Modularity in the Syntactic Parser. In: Garfield, J.L. (ed.) Modularity in Knowledge Representation and Natural-Language Understanding, pp. 259–276. MIT Press, Cambridge (1987)
A 3D Conversational Agent for Presenting Digital Information for Deaf People Da Phuc Phan, Thi Nhat Thanh Nguyen, and The Duy Bui College of Technology Vietnam National University, Hanoi [email protected]
Abstract. Sign languages appear to be the major communication tool for the deaf community. As a minor community, Vietnamese deaf people often have to rely on a small numbers of interpreters who are hearing people knowing sign language in order to communicate with hearing people. It would be very beneficial for deaf as well as hearing people to develop an automated translation system between spoken/written languages and sign language. In this paper, we present our attempt to graphically decompose the gestures in Vietnamese Sign Language so that they can be easily synthesized in a 3D conversational agent. In order to describe the signs in Vietnamese Sign Language, we proposed an animation-level markup language. Based on this, we presented our 3D conversational agent for presenting multi-modal information for deaf people. The 3D agent can assist deaf people easier and in a more natural way. Keywords: Sign Language Synthesis, Human Computer Interaction, 3D Conversational Agents.
1
Introduction
Sign languages have the similar communication capacities to those of spoken languages. They appear to be the major communication tool for the deaf community [6,9]. Sign languages include finger spelling and gesture. In finger spelling, finger postures and motions are used to express letters. Through a letter sequence, a sign word is spelled. On the other hand, gestures and motions are used to express a word, or even a phrase. Each gesture is described by hand shape, location, movement, orientation, number of hands. Sign language has its own grammar based on spatial, physiological and temporal dimensions. For example, not only strongly recognized gestures, but also pronouns and index references, appeal to mechanisms involving the signers space. Beside finger spelling and gesture, other movements of the body, including mouth patterns, head and shoulder movements and facial expression are used in sign languages to express speaker’s thoughts. These movements both play paralinguistic (kinetic stress, face expression) and grammatical roles (marking syntactic clauses). In Vietnam, there are about 5.3 million disabled people, a large part of that are deaf people. As a minor community, deaf people often experience great difficulties A. Ghose, G. Governatori, and R. Sadananda (Eds.): PRIMA 2007, LNAI 5044, pp. 319–328, 2009. c Springer-Verlag Berlin Heidelberg 2009
320
D.P. Phan, T.N.T. Nguyen, and T.D. Bui
in communicating with hearing people. They have to rely on a small number of interpreters who are hearing people knowing sign language. As the society is paying more and more attention to deaf people in recent years, efforts have been made to produce tools that allow them to communicate easier with hearing people. Fortunately, recent advanced research in linguistics, machine learning and computer graphics have offered different models for automated translation between spoken/written languages and sign language. It would be very beneficial for deaf as well as hearing people to develop such a translation system. As is the case in spoken language, Vietnamese Sign Language differs from one region to another. As in the past, gestures and signs of the deaf in Vietnam have developed the best in such places as Hanoi, Haiphong, Hochiminh City,... These gestures and signs are often natural, spontaneous or imported from overseas and have not been systematized yet. In order to help the deaf have the opportunity to select the signs they wish to learn, a project has been deployed by University of Pedagogy, Ho Chi Minh City to develope a “Vietnamese Sign Language Dictionary For People With Hearing Impairment” [12], in which the signs of three aforementioned cities and the Common Signs have been recorded. The dictionary comprises popular notion in topics related to daily life to help the hearing and deaf people communicate and study more effectively. In Vietnamese Sign Language, however, there have not been either any serious attempt of decoding of the language structure or synthesizing the language for automated translation. In this paper, we attempt to graphically decompose the gestures in Vietnamese Sign Language so that they can be easily synthesized in a 3D avatar. We propose an animation-level markup language to describe the signs in Vietnamese Sign Language. Based on this, we propose a method of presenting multi-modal information for deaf people with the use of a 3D conversational agent. The 3D agent is built based on Java3D, VRML (Virtual Reality Modeling Language) and H-anim, an abstract representation for modeling three dimensional human figures. The rest of the paper is organized as follows. Section 2 summarizes some background concepts on Sign Language and notations to represent Sign Language. We then review some related work in Section 3. In Section 4, we present our markup language for representing Vietnamese Sign Language. Section 5 shows how we present information for deaf people in a 3D agent.
2
Background
Following Stokoes footsteps [10], linguists have proved that sign languages, like oral ones, were articulated into phonemic and morphemic levels. At phonemic level, the basic movements - phonemes, which are formerly called cheremes, are functionally equivalent to the phonemes of spoken languages. It has been recognized that for phonemes in sign languages, the mental abstractions involved are essentially the same as in oral languages. Phonemes in a sign language are usually comprised of: hand shape, location, movement, and palm orientation. An individual hand shape means nothing out of context. It is only when we combine
A 3D Conversational Agent for Presenting Digital Information
321
that hand shape with movement and a certain location that we able to determine its meaning. At morphemic level, a morpheme is the smallest linguistic unit that has semantic meaning. Similar to spoken language, morphemes are composed of phonemes. Sign languages are often represented in some visual forms for the purpose of communication and education. There are three popular ways to represent sign languages, which are Stokoe Notation [10], Sutton SignWriting [2], and HamNoSys [8]. The Stokoe notation [10] for American Sign Language was created earliest for a sign language and is mostly restricted to linguists and academics. It it is arranged linearly on the page and uses elements of the Latin alphabet and is phonemic. This system only uses a reduced set of symbols to match the needs of American Sign Language. For example, only a single symbol is used for circling movement, regardless of whether the plane of the movement is horizontal or vertical because the other parameter of the motion is determined by American Sign Language phonotactics and need not be indicated in a phonemic system. Sutton SignWriting [2], created by Valerie Sutton in 1974, represent the gesture and movements of signed languages by visual symbols. It is based on a notation system for dance movements representation - Sutton DanceWriting. As a “movement-writing-alphabet”, SignWriting can be used to describe any signed language in the world. The Hamburg Sign Language Notation System, HamNoSys [8], was created in 1989 by a group of deaf and hearing people. Unlike SignWriting, HamNoSys was developed for study and research purpose. HamNoSys breaks each sign down into about 200 characters in some components such as hand position, hand orientation, hand shape, motion, etc. All of signs are figure shape so it is easy to read by human.
3
Related Work
There have been a number of attempts in synthesizing sign languages. In SignAnim project (also known as Simon-the-Signer) [5,7,13], a signing avatar system based on motion capture is used. Each sign is represented by a data file recording motion parameters for the body, arms, hands, and so on. The corresponding sequence of such motion data files can be used to animate the 3D humanoid. In this way, accurate motions can be created. However, the major drawback of motion capture is that we have to record the large number of signs for a complete lexicon. To overcome the drawback of motion capture, some projects synthesize deaf signing animation from a high-level description of signs in terms of the HamNoSys transcription system [8]. This approach use the HamNoSys notation for transcribing signing gestures. This notation system breaks each sign down into components such as hand position, hand orientation, hand shape, motion, etc. Using HamNoSys notation can improve the disadvantage of motion capture. However, this notation is designed to be read by humans rather than computers. Therefore it is quite difficult to implement. Adding a new sign into this notation is also not easy. An example of this approach is the ViSiCAST project [1]. In this project, an XML-compatible notation, ViSiCAST-GML, is developed based on
322
D.P. Phan, T.N.T. Nguyen, and T.D. Bui
HamNoSys [8] in order to produce adaptable communication tools allowing sign language communication where only speech and text are available. In Thetos project [11], input data in form of a text file in Polish is linguistically analyzed to produce the output in form of an animated sequence. All the existing work for sign language synthesizing, however, did not concentrate on creating an easy method for mapping sign language to animation sequence.
4
Animation Markup Language for Vietnamese Sign Language
Vietnamese Sign Language is based on the well-established American Sign Language. Thus, there are about more than 600 concepts in Vietnamese Sign Language, which can be decomposed into morphemes and phonemes [4]. For other concepts, it is possible to implement finger spelling of Vietnamese words similar to the American Sign Language system. Figure 1 shows the alphabet table of Vietnamese Sign Language. Each sign/morpheme in Vietnamese Sign Language maybe a static shape or a dynamic gesture. A gesture starts with a static shape followed by a hand movement. In general, a sign /morpheme can be articulated into phonemes, which are: hand shape, palm orientation, hand position and hand movement.
Fig. 1. The alphabet table of Vietnamese Sign Language
4.1
Hand Shape
There are many different types of hand shape in Vietnamese Sign Language. Some typical hand shapes are: hold shape, large open shape, small open shape, C shape (hand like C character), finger point shape and hook shape. Visibly, the hand shapes can be distinguished by the level of bending of each finger and the distance between fingers. We propose 5 levels of bending, ranging from very
A 3D Conversational Agent for Presenting Digital Information
323
small to very big. These 5 levels are enough to describe signs in Vietnamese. The distance between each finger is small or big. The hand shape can also be used to describe postures in finger spelling. Here are how we describe a hand shape in XML: very small/ small/ medium/ big/ very big very small/ small/ medium/ big/ very big <middle> very small/ small/ medium/ big/ very big very small/ small/ medium/ big/ very big very small/ small/ medium/ big/ very big small/big
where , , <middle>, , are the level of bending of thumb, index finger, middle finger, ring finger and pinky respectively; represents the distance between fingers. 4.2
Palm Orientation
The palm can be in some main orientations: inside, outside, up, down, left and right. Here are how we describe palm orientation in XML: <palm orientation> <X>0/-1/1 0/-1/10/-1/1
where the palm orientation is determined by three X, Y, Z values, with 0 meaning orientation independent, 1 meaning inside (for Z)/up (for Y)/left (for X), and -1 meaning outside (for Z)/down (for Y)/right (for X). 4.3
Hand Position
In Vietnamese Sign Language, the hand can be set into many position to represent different meanings. For example, pointing index finger to the nose describes the concept “smell”. The position of the hand can be determined with regard to: head, left side of the head (l head), right side of the head (r head), chest, left side of the chest (l chest), right side of the chest (r chest), waist, eye, nose and mouth. Moreover, the dependency between two hands, e.g. left hand above right hand, is also a part of hand position. Here are how we describe hand position in XML: dependence/independence <depend post> up/down/equal <post>head/l head/r head/chest/l chest/r chest/waist/nose/eye/mouth
where indicates the dependency between two hands; <depend post> indicates the type of dependency between two hands; and <post> indicates the position of the hand with regard to the body.
324
4.4
D.P. Phan, T.N.T. Nguyen, and T.D. Bui
Movement
Movements in sign language can be divided into 2 different categories: finger movement and hand movement. Finger movement can be of several form: wave form, the thumb and index finger touch together, and so on. Finger movement can also be repeated several times. Here are how we describe finger movement in XML: <finger move> wave/thumb index/thumb pinky/open/close/shake
where contains the form of finger movement, which can be wave, thumb index (the thumb and index finger touch each other and move in reverse directions), thumb pinky (the thumb in turn touches from index finger tip to pinky tip, and move in reverse directions), open (hand in closed position, then opening fingers out), close and shake;
